1 00:00:00,000 --> 00:00:03,940 So back in 1985, an industrial robot on a factory 2 00:00:03,940 --> 00:00:07,480 floor just, you know, suddenly accelerated in 3 00:00:07,480 --> 00:00:09,779 a direction it mathematically was not supposed 4 00:00:09,779 --> 00:00:11,980 to move. Right. Which is absolutely terrifying 5 00:00:11,980 --> 00:00:13,759 when you're standing next to it. Yeah. I mean, 6 00:00:13,759 --> 00:00:16,359 it nearly took a worker's head off. And the crazy 7 00:00:16,359 --> 00:00:19,359 part is the motors weren't broken. Nope. Hardware 8 00:00:19,359 --> 00:00:22,239 was fine. The code didn't even have a typo. The 9 00:00:22,239 --> 00:00:25,399 robot had just hit something called a singularity. 10 00:00:25,539 --> 00:00:28,059 Right. A geometric black hole. Exactly. A black 11 00:00:28,059 --> 00:00:29,760 hole where the laws of robotics basically just 12 00:00:29,899 --> 00:00:33,600 divide by zero. So, welcome to the deep dive. 13 00:00:33,859 --> 00:00:37,420 It's a phenomenal topic today. I mean, that singularity 14 00:00:37,420 --> 00:00:39,439 problem is something roboticists still wrestle 15 00:00:39,439 --> 00:00:41,880 with. Every single time they plan a trajectory. 16 00:00:42,140 --> 00:00:44,100 Absolutely. Yeah, and if you're designing the 17 00:00:44,100 --> 00:00:46,539 machines that build our world or, you know, if 18 00:00:46,539 --> 00:00:48,460 you just want to understand the terrifyingly 19 00:00:48,460 --> 00:00:50,700 beautiful math that keeps them from tearing themselves 20 00:00:50,700 --> 00:00:53,119 apart, you are in the right place. Definitely. 21 00:00:53,320 --> 00:00:56,460 So we are continuing our Marginalia series, diving 22 00:00:56,460 --> 00:00:59,520 back into Peter Corke's definitive 2023 textbook, 23 00:01:00,020 --> 00:01:02,679 Robotics, Vision, and Control, Fundamental Algorithms 24 00:01:02,679 --> 00:01:05,340 from Python. A true classic. Yeah. And this is 25 00:01:05,340 --> 00:01:08,099 part three of our five -part journey. Just to 26 00:01:08,099 --> 00:01:10,159 bridge the gap for a second, if you think back 27 00:01:10,159 --> 00:01:12,280 to part two, we were out in the world, right? 28 00:01:12,280 --> 00:01:14,140 Right. Right, we were looking at mobile robots 29 00:01:14,140 --> 00:01:16,379 navigating through these unstructured environments. 30 00:01:16,599 --> 00:01:18,519 Doing SLAM, figuring out where they were. The 31 00:01:18,519 --> 00:01:22,299 core problem was localization, but today we are 32 00:01:22,299 --> 00:01:24,239 taking the robot and we are bolting it to the 33 00:01:24,239 --> 00:01:26,680 floor. We are entirely shifting the mathematical 34 00:01:26,680 --> 00:01:29,040 framework. I mean, when you anchor the base, 35 00:01:29,500 --> 00:01:31,560 localization ceases to be the primary problem. 36 00:01:32,040 --> 00:01:35,450 Because the base frame is completely fixed. Exactly. 37 00:01:35,769 --> 00:01:37,709 The challenge now becomes pure manipulation. 38 00:01:38,250 --> 00:01:40,250 You aren't moving through a space anymore. You 39 00:01:40,250 --> 00:01:42,890 are reaching into a highly constrained, specific 40 00:01:42,890 --> 00:01:46,049 workspace. Just physically altering it. Right. 41 00:01:46,049 --> 00:01:48,569 And you need extreme precision. And, well, eventually 42 00:01:48,569 --> 00:01:51,849 you need to exert calculated forces. So we are 43 00:01:51,849 --> 00:01:54,250 exploring chapters seven, eight, and nine of 44 00:01:54,250 --> 00:01:56,829 Cork's text today. And we have this very specific 45 00:01:56,829 --> 00:01:59,670 mission, this evolution to trace. It's quite 46 00:01:59,670 --> 00:02:02,920 a journey. It really is. We start with the pure, 47 00:02:03,480 --> 00:02:06,280 rigid geometry of a robotic arm just frozen in 48 00:02:06,280 --> 00:02:08,879 time. The skeleton, basically. Yeah. And once 49 00:02:08,879 --> 00:02:12,199 we have that skeleton mapped, we inject differential 50 00:02:12,199 --> 00:02:14,979 calculus to understand how it moves. The velocity. 51 00:02:15,319 --> 00:02:18,580 Right. And finally, we just crash headfirst into 52 00:02:18,580 --> 00:02:21,500 the messy, unforgiving reality of physics, like 53 00:02:21,500 --> 00:02:24,439 mass, inertia, and gravity. It's the only logical 54 00:02:24,439 --> 00:02:26,400 way to build up the architecture of a manipulator, 55 00:02:26,620 --> 00:02:28,800 honestly. Because you can't just skip straight 56 00:02:28,800 --> 00:02:31,300 to the forces, right? No, you cannot reliably 57 00:02:31,300 --> 00:02:33,719 control the torque a robot applies to a piece 58 00:02:33,719 --> 00:02:36,639 of steel until you understand the velocity vectors 59 00:02:36,639 --> 00:02:40,039 at its tooltip. Makes sense. And you can't understand 60 00:02:40,039 --> 00:02:42,800 those velocities until you have an airtight model 61 00:02:42,800 --> 00:02:45,759 of the underlying joint geometry. Cork builds 62 00:02:45,759 --> 00:02:48,620 this foundation beautifully. OK, so let's start 63 00:02:48,620 --> 00:02:51,020 with that geometric skeleton. Before we can even 64 00:02:51,020 --> 00:02:53,080 look at the math, we have to look at the mechanical 65 00:02:53,080 --> 00:02:55,599 reality of the hardware itself. Right. What is 66 00:02:55,599 --> 00:02:58,819 the arm actually made of? Exactly. Cork defines 67 00:02:58,819 --> 00:03:01,479 a robot manipulator arm as a series of rigid 68 00:03:01,479 --> 00:03:03,840 links connected by computer controllable joints. 69 00:03:04,479 --> 00:03:06,900 But right away he draws a really hard line between 70 00:03:06,900 --> 00:03:09,120 the serial link manipulators and parallel link 71 00:03:09,120 --> 00:03:11,740 manipulators. It's a fundamental divide. And 72 00:03:11,740 --> 00:03:14,139 for anyone who has ever tried to design, say, 73 00:03:14,280 --> 00:03:16,139 a high -speed pick -and -place machine, this 74 00:03:16,139 --> 00:03:18,419 is a hardware debate that dictates your entire 75 00:03:18,419 --> 00:03:20,800 software stack. It absolutely dictates the math. 76 00:03:21,000 --> 00:03:22,620 I mean, when most people picture an industrial 77 00:03:22,620 --> 00:03:25,580 robot, they visualize a serial link manipulator. 78 00:03:25,900 --> 00:03:28,500 The classic six degree of freedom factory arm. 79 00:03:28,800 --> 00:03:32,240 Yeah, like the old Puma 560 or a modern KUKA 80 00:03:32,240 --> 00:03:35,960 arm. They are anthropomorphic. Like a human arm. 81 00:03:36,319 --> 00:03:39,199 Exactly. You have a base, a shoulder, an upper 82 00:03:39,199 --> 00:03:42,180 arm, an elbow, a forearm, a wrist, and an end 83 00:03:42,180 --> 00:03:45,560 defector. They're connected in series. It's a 84 00:03:45,560 --> 00:03:48,039 kinematic chain. It is a kinematic chain. Right, 85 00:03:48,139 --> 00:03:50,659 a chain of bones. But the mechanical consequence 86 00:03:50,659 --> 00:03:53,520 of that serial chain is brutal from a physics 87 00:03:53,520 --> 00:03:55,539 perspective, isn't it? Oh, completely brutal. 88 00:03:55,780 --> 00:03:57,979 Because every single joint has to support the 89 00:03:57,979 --> 00:03:59,659 weight, the inertia, and the movement of every 90 00:03:59,659 --> 00:04:01,500 single joint that comes after it. Precisely. 91 00:04:01,780 --> 00:04:03,719 The shoulder joint isn't just moving the arm. 92 00:04:04,039 --> 00:04:06,860 It is carrying the elbow motor. And the wrist 93 00:04:06,860 --> 00:04:09,659 motors. Right, and the heavy metal links, the 94 00:04:09,659 --> 00:04:11,599 gripper, and whatever payload you are holding. 95 00:04:11,939 --> 00:04:15,139 That's a lot of weight. It is. This compounding 96 00:04:15,139 --> 00:04:17,399 mass means the motors at the base have to be 97 00:04:17,399 --> 00:04:19,920 absolutely massive to generate the required torque. 98 00:04:20,339 --> 00:04:22,699 Just to lift itself. Exactly. And because the 99 00:04:22,699 --> 00:04:25,480 arm itself is so heavy, its structural rigidity 100 00:04:25,480 --> 00:04:27,639 is constantly compromised by its own weight. 101 00:04:27,939 --> 00:04:30,399 Which brings us to the alternative, right? Parallel 102 00:04:30,399 --> 00:04:33,339 link manipulators? If I'm building a Delta robot 103 00:04:33,339 --> 00:04:35,519 to sort batteries on a high -speed conveyor belt, 104 00:04:35,939 --> 00:04:38,079 a serial arm is just going to tear itself apart 105 00:04:38,079 --> 00:04:40,199 from the inertia. It would vibrate to pieces. 106 00:04:40,379 --> 00:04:42,759 In a parallel robot like that Delta configuration 107 00:04:42,759 --> 00:04:45,029 you mentioned, You remove the motors from the 108 00:04:45,029 --> 00:04:47,970 moving structure entirely. Oh, wow. All the heavy 109 00:04:47,970 --> 00:04:50,470 actuators are bolted stationary to the base frame. 110 00:04:50,750 --> 00:04:53,329 So they don't move at all. Right. They just push 111 00:04:53,329 --> 00:04:55,550 and pull on these lightweight carbon fiber or 112 00:04:55,550 --> 00:04:58,170 aluminum linkages that all meet down at the end 113 00:04:58,170 --> 00:05:01,149 effector in parallel. So the tool is supported 114 00:05:01,149 --> 00:05:03,509 by multiple independent chains simultaneously. 115 00:05:03,910 --> 00:05:06,829 Yes, exactly. The moving mass is incredibly low, 116 00:05:06,990 --> 00:05:09,629 which allows for astonishingly high accelerations. 117 00:05:10,069 --> 00:05:12,170 I've seen videos of those things. They are a 118 00:05:12,170 --> 00:05:14,009 blur. They look like they're breaking the laws 119 00:05:14,009 --> 00:05:17,230 of physics. But the trade -off is the workspace. 120 00:05:17,569 --> 00:05:20,730 Ah, right. They can't reach very far. No. A serial 121 00:05:20,730 --> 00:05:22,850 arm can reach around obstacles. It can reach 122 00:05:22,850 --> 00:05:26,250 behind itself. A parallel robot usually has this 123 00:05:26,250 --> 00:05:28,430 very limited dome -shaped workspace directly 124 00:05:28,430 --> 00:05:30,750 beneath it. And the math is totally different. 125 00:05:31,189 --> 00:05:33,790 Oh, the kinematic equations to control it are 126 00:05:33,790 --> 00:05:36,189 completely reversed in difficulty, which we'll 127 00:05:36,189 --> 00:05:39,120 get into. Well, for the vast majority of Cork's 128 00:05:39,120 --> 00:05:41,680 focus here and our focus today, we are looking 129 00:05:41,680 --> 00:05:44,819 at those serial link arms, the absolute workhorses 130 00:05:44,819 --> 00:05:46,740 of the automotive and manufacturing industries. 131 00:05:46,920 --> 00:05:49,720 Right. The classic chain. Now, the bones of these 132 00:05:49,720 --> 00:05:52,259 arms are connected by joints. And while mechanical 133 00:05:52,259 --> 00:05:54,959 engineers have like a million different complex 134 00:05:54,959 --> 00:05:57,620 joint types, Cork boils the robotic kinematic 135 00:05:57,620 --> 00:06:00,480 model down to two primary primitives. Just two. 136 00:06:00,699 --> 00:06:03,930 Yeah. Revolut joints and prismatic joints. Because 137 00:06:03,930 --> 00:06:05,569 mathematically you can build almost anything 138 00:06:05,569 --> 00:06:07,529 out of those two. Okay, break them down for me. 139 00:06:08,149 --> 00:06:10,829 A revolute joint is a pure rotational degree 140 00:06:10,829 --> 00:06:13,209 of freedom. Basically a hinge. Like an elbow? 141 00:06:13,670 --> 00:06:16,850 Exactly. One link rotates relative to the previous 142 00:06:16,850 --> 00:06:20,209 one around a single defined axis. Got it. And 143 00:06:20,209 --> 00:06:23,829 prismatic. A prismatic joint is a pure translational 144 00:06:23,829 --> 00:06:26,810 degree of freedom. Think of a linear slider or 145 00:06:26,810 --> 00:06:29,750 a telescopic boom. So it slides in and out. Right, 146 00:06:29,769 --> 00:06:32,329 it extends or retracts along an axis without 147 00:06:32,329 --> 00:06:35,000 any rotation. Cork brings up the Stanford arm 148 00:06:35,000 --> 00:06:36,959 here, which I think is a brilliant historical 149 00:06:36,959 --> 00:06:39,220 anchor. Victor Scheinman developed it back in 150 00:06:39,220 --> 00:06:42,379 1969. It is a seminal piece of hardware, truly. 151 00:06:42,660 --> 00:06:44,720 And it's a fascinating hybrid, right? Because 152 00:06:44,720 --> 00:06:47,000 it utilizes a really prominent prismatic joint. 153 00:06:47,160 --> 00:06:49,560 It does. The first two joints of the Stanford 154 00:06:49,560 --> 00:06:52,540 arm are revolute, so they create a two -axis 155 00:06:52,540 --> 00:06:55,360 shoulder that can point in any direction. Like 156 00:06:55,360 --> 00:06:58,519 a turret. Yeah, like a turret. But the third 157 00:06:58,519 --> 00:07:01,319 joint is prismatic. It's a boom that telescopes 158 00:07:01,319 --> 00:07:03,920 in and out. So instead of an elbow that bends, 159 00:07:03,959 --> 00:07:06,639 you have an arm that extends. Exactly. Scheinman 160 00:07:06,639 --> 00:07:09,439 actually used the insights from building that 161 00:07:09,439 --> 00:07:12,480 arm to later design the Puma robots for unimation. 162 00:07:12,759 --> 00:07:16,040 Oh, that's right. Cork notes that the Puma 560 163 00:07:16,040 --> 00:07:18,660 became the absolute, quote unquote, white rat 164 00:07:18,660 --> 00:07:21,220 of robotics research for decades. It really was. 165 00:07:21,680 --> 00:07:24,139 If you wrote a new control algorithm in the 80s 166 00:07:24,139 --> 00:07:28,420 or 90s, you proved it on the Puma 560. OK, so 167 00:07:28,420 --> 00:07:30,740 we have our links and we have a revolute or prismatic 168 00:07:30,740 --> 00:07:33,680 joints. The ultimate objective is to place the 169 00:07:33,680 --> 00:07:35,939 end effector, you know, the welding torch, the 170 00:07:35,939 --> 00:07:38,879 gripper, the camera, in a precise pose in 3D 171 00:07:38,879 --> 00:07:42,779 space. Position and orientation. Right. And that 172 00:07:42,779 --> 00:07:44,720 requires us to solve the fundamental problem 173 00:07:44,720 --> 00:07:47,839 of forward kinematics. Forward kinematics is 174 00:07:47,839 --> 00:07:50,060 the definitive deterministic mapping from the 175 00:07:50,060 --> 00:07:52,899 joint space to the task space. Let's define those 176 00:07:52,899 --> 00:07:56,069 spaces really quick. Sure. If I query the encoders 177 00:07:56,069 --> 00:07:58,829 and you tell me exactly how many radians every 178 00:07:58,829 --> 00:08:02,029 revolute joint is rotated and how many meters 179 00:08:02,029 --> 00:08:05,490 every prismatic joint is extended, that's the 180 00:08:05,490 --> 00:08:08,610 joint space. The raw motor data. Exactly. Forward 181 00:08:08,610 --> 00:08:11,589 Kinematics will take that data and output the 182 00:08:11,589 --> 00:08:14,170 exact SE3 transformation matrix representing 183 00:08:14,170 --> 00:08:17,850 the position and orientation of the tooltip relative 184 00:08:17,850 --> 00:08:21,459 to the base. That's the task space. To do this, 185 00:08:21,600 --> 00:08:23,319 cork leverage is a concept we actually touched 186 00:08:23,319 --> 00:08:26,800 on before in part one, compounding relative poses. 187 00:08:27,160 --> 00:08:30,319 Yes, the pose graphs. Right. You don't try to 188 00:08:30,319 --> 00:08:32,419 calculate the hands position relative to the 189 00:08:32,419 --> 00:08:35,779 base in one giant messy equation. That'd be impossible. 190 00:08:35,940 --> 00:08:38,679 You just break it down link by link. It is entirely 191 00:08:38,679 --> 00:08:41,340 recursive. You establish a coordinate frame at 192 00:08:41,340 --> 00:08:43,179 the base of the robot. Let's call that frame 193 00:08:43,179 --> 00:08:46,120 zero. Okay, the anchor. Then you establish frame 194 00:08:46,120 --> 00:08:49,029 one at the very first joint. The geometric relationship 195 00:08:49,029 --> 00:08:51,929 between frame zero and frame one is defined by 196 00:08:51,929 --> 00:08:54,370 a homogeneous transformation matrix. A four by 197 00:08:54,370 --> 00:08:57,149 four matrix. Right. That matrix contains the 198 00:08:57,149 --> 00:08:59,210 fixed mechanical offsets of the robot's base 199 00:08:59,210 --> 00:09:01,809 plus the variable rotation of that first motor. 200 00:09:02,190 --> 00:09:04,009 Then you just chain them together. Frame one 201 00:09:04,009 --> 00:09:06,509 to frame two, frame two to frame three. You mathematically 202 00:09:06,509 --> 00:09:09,429 compound them using matrix multiplication. It's 203 00:09:09,429 --> 00:09:12,299 just a chain of multiplication. Exactly. In Cork's 204 00:09:12,299 --> 00:09:14,860 robotics toolbox for Python, he implements this 205 00:09:14,860 --> 00:09:17,139 using an elementary transformation sequence, 206 00:09:17,480 --> 00:09:20,559 or ETS. ETS, right. Which feels so intuitive. 207 00:09:21,000 --> 00:09:23,980 It is incredibly intuitive. You are literally 208 00:09:23,980 --> 00:09:26,500 just multiplying a sequence of elementary translations 209 00:09:26,500 --> 00:09:28,779 and rotations. Like step -by -step instructions. 210 00:09:29,039 --> 00:09:32,159 Yes, like translate along x by the link length, 211 00:09:32,639 --> 00:09:35,139 rotate around z by the joint angle, translate 212 00:09:35,139 --> 00:09:37,600 along z by the offset. You just keep doing that. 213 00:09:37,919 --> 00:09:40,200 You string 15 or 20 of these elementary matrices 214 00:09:40,200 --> 00:09:43,000 together, and when you plug in the current joint 215 00:09:43,000 --> 00:09:44,840 variables, the whole sequence multiplies out. 216 00:09:45,159 --> 00:09:48,200 Just collapsing into a single 4x4 matrix. Exactly. 217 00:09:48,340 --> 00:09:51,820 That. SE3 matrix. The top left 3x3 block gives 218 00:09:51,820 --> 00:09:54,519 you the rotation, the orientation of the tool, 219 00:09:55,340 --> 00:09:58,440 and the top right 3x1 vector gives you the XYZ 220 00:09:58,440 --> 00:10:01,039 translation. It's clean, it's computationally 221 00:10:01,039 --> 00:10:03,379 fast, and it perfectly describes the static posture 222 00:10:03,379 --> 00:10:06,179 of the robot. It is clean. when you use an ETS 223 00:10:06,179 --> 00:10:11,649 string, but historically, robotics had to standardize 224 00:10:11,649 --> 00:10:14,470 how these kinematic chains were published in 225 00:10:14,470 --> 00:10:16,669 research papers and manufacturer spec sheets. 226 00:10:16,870 --> 00:10:19,629 Uh -oh. I feel a headache coming on. What's fascinating 227 00:10:19,629 --> 00:10:22,870 here is that this is where Cork dives into the 228 00:10:22,870 --> 00:10:25,970 absolute headache that is Denevet -Hartenberg 229 00:10:25,970 --> 00:10:29,909 parameters. Oh, man, the DH parameters. I mean, 230 00:10:29,990 --> 00:10:32,190 if you look at modern ROS implementations, everyone 231 00:10:32,190 --> 00:10:35,110 just uses URDF, right? The Unified Robot Description 232 00:10:35,110 --> 00:10:37,870 Format. Yeah, URDF is the standard now for software. 233 00:10:38,409 --> 00:10:40,909 Because it's just an XML file. It explicitly 234 00:10:40,909 --> 00:10:43,809 lists the origin of every joint, the axis of 235 00:10:43,809 --> 00:10:47,070 rotation, the visual meshes, the inertia tensors. 236 00:10:47,169 --> 00:10:49,710 It's highly readable and completely unambiguous. 237 00:10:49,929 --> 00:10:52,970 So why do we care about DH, right? Exactly. Why 238 00:10:52,970 --> 00:10:55,289 hasn't URDF just entirely replaced them in the 239 00:10:55,289 --> 00:10:57,870 textbook? That is a very common frustration for 240 00:10:57,870 --> 00:11:01,149 engineers entering the field today. URDF is fantastic 241 00:11:01,149 --> 00:11:03,590 for simulation and for software stacks like ROS. 242 00:11:03,990 --> 00:11:07,610 Right. But URDF is bloated. It contains a massive 243 00:11:07,610 --> 00:11:09,389 amount of data that is completely irrelevant 244 00:11:09,389 --> 00:11:12,250 to the pure kinematic algebra. Like the visual 245 00:11:12,250 --> 00:11:14,409 mesh of the robot arm? The math doesn't care 246 00:11:14,409 --> 00:11:17,950 what it looks like? Exactly. published their 247 00:11:17,950 --> 00:11:21,610 convention back in 1955 as a minimal mathematical 248 00:11:21,610 --> 00:11:25,710 representation. Minimal being the key word. They 249 00:11:25,710 --> 00:11:27,870 proved that you can describe the geometric relationship 250 00:11:27,870 --> 00:11:30,789 between any two adjacent links using only four 251 00:11:30,789 --> 00:11:34,490 parameters. Just four. Four numbers. Link length, 252 00:11:34,970 --> 00:11:38,870 link twist, link offset, and joint angle. So 253 00:11:38,870 --> 00:11:42,629 a six axis robot is entirely described by a table 254 00:11:42,629 --> 00:11:46,669 of just 24 numbers. Exactly. It boils the entire 255 00:11:46,669 --> 00:11:49,970 3D geometry down into an ultra -compact matrix 256 00:11:49,970 --> 00:11:53,110 of four columns. Okay, I see the appeal. And 257 00:11:53,110 --> 00:11:55,129 the reason it survives is that it forces the 258 00:11:55,129 --> 00:11:57,889 coordinate frames to be assigned in a very specific, 259 00:11:58,350 --> 00:12:00,710 mathematically rigorous way. Like strict rules 260 00:12:00,710 --> 00:12:04,289 on where the X and Y axes go? Yes. The X axis 261 00:12:04,289 --> 00:12:07,230 of a frame must intersect the Z axis of the previous 262 00:12:07,230 --> 00:12:10,309 frame, and it must be perpendicular to it. And 263 00:12:10,309 --> 00:12:13,169 when you enforce those constraints... a lot of 264 00:12:13,169 --> 00:12:15,950 terms in the transformation matrices become zero. 265 00:12:16,580 --> 00:12:19,980 Oh, so it simplifies the algebra. Immensely. 266 00:12:20,360 --> 00:12:22,940 If you were trying to derive closed -form algebraic 267 00:12:22,940 --> 00:12:24,820 solutions for a robot, which we will get to in 268 00:12:24,820 --> 00:12:27,399 a minute, having all those zeros in your matrices 269 00:12:27,399 --> 00:12:30,480 makes the symbolic math manageable. You're not 270 00:12:30,480 --> 00:12:32,840 multiplying a bunch of random decimals. Right. 271 00:12:33,139 --> 00:12:35,960 You can't easily do symbolic algebraic manipulation 272 00:12:35,960 --> 00:12:40,059 on a bloated URDF file. DH parameters are the 273 00:12:40,059 --> 00:12:43,779 language of kinematic proofs. Okay, I concede 274 00:12:43,779 --> 00:12:47,240 the mathematical elegance. Right. But Quark explicitly 275 00:12:47,240 --> 00:12:49,799 calls out that DH parameters are the source of 276 00:12:49,799 --> 00:12:52,460 catastrophic confusion in practice. Oh, they 277 00:12:52,460 --> 00:12:55,019 are a nightmare in practice. Because there isn't 278 00:12:55,019 --> 00:12:57,980 just one DH convention, is there? This is the 279 00:12:57,980 --> 00:13:00,980 nightmare scenario. There are two competing conventions 280 00:13:00,980 --> 00:13:03,480 that are both widely used in literature. Why 281 00:13:03,480 --> 00:13:05,850 do we do this to ourselves? I don't know. but 282 00:13:05,850 --> 00:13:08,870 we have standard DH parameters, which follow 283 00:13:08,870 --> 00:13:11,429 Denavit and the Hartenberg's original 1955 paper, 284 00:13:11,429 --> 00:13:14,629 and we have modified DH parameters. Where did 285 00:13:14,629 --> 00:13:17,169 the modified ones come from? They were popularized 286 00:13:17,169 --> 00:13:19,590 by John Craig in his famous robotics textbook 287 00:13:19,590 --> 00:13:23,850 in the 1980s. Okay, let me guess. They look identical 288 00:13:23,850 --> 00:13:27,490 in a table, but the underlying geometry is fundamentally 289 00:13:27,490 --> 00:13:30,090 different. They look exactly identical in a table. 290 00:13:30,190 --> 00:13:32,330 It is literally just a matrix of four columns. 291 00:13:32,809 --> 00:13:37,279 A, alpha, D, Theta. But the math is different. 292 00:13:37,799 --> 00:13:40,159 The transformation sequence is completely different. 293 00:13:41,039 --> 00:13:44,100 In standard DH, the parameters describe the transformation 294 00:13:44,100 --> 00:13:46,539 from the current link to the next link. OK, so 295 00:13:46,539 --> 00:13:48,440 looking forward. Right. The frames are attached 296 00:13:48,440 --> 00:13:51,100 to the distal end of the links. But in modified 297 00:13:51,100 --> 00:13:53,440 DH, the parameters describe the transformation 298 00:13:53,440 --> 00:13:55,840 from the previous link to the current one. Looking 299 00:13:55,840 --> 00:13:58,320 backward. Yes. The frames are attached to the 300 00:13:58,320 --> 00:14:02,279 proximal end. So if I buy a KUKA arm and the 301 00:14:02,279 --> 00:14:05,299 manufacturer provides a datasheet with a DH table, 302 00:14:05,580 --> 00:14:07,679 and I just blindly type those 24 numbers into 303 00:14:07,679 --> 00:14:10,340 a Python script. If the datasheet used standard 304 00:14:10,340 --> 00:14:13,720 DH and your Python algorithm or your C++ solver 305 00:14:13,720 --> 00:14:16,980 was written expecting modified DH, the resulting 306 00:14:16,980 --> 00:14:19,500 kinematic model is garbage. Just totally wrong. 307 00:14:19,759 --> 00:14:21,620 It isn't just a little bit off. The coordinate 308 00:14:21,620 --> 00:14:23,820 frames are physically in the wrong places in 309 00:14:23,820 --> 00:14:26,360 the simulation. Oh, man. If you command the robot 310 00:14:26,360 --> 00:14:28,789 to move based on that model, The simulation might 311 00:14:28,789 --> 00:14:31,629 show the arm moving gracefully up, while the 312 00:14:31,629 --> 00:14:33,769 physical thousand -pound arm swings down and 313 00:14:33,769 --> 00:14:37,070 crushes a workbench. That is terrifying. It means 314 00:14:37,070 --> 00:14:39,950 a naked table of DH parameters is functionally 315 00:14:39,950 --> 00:14:41,909 useless unless you explicitly know the convention 316 00:14:41,909 --> 00:14:45,570 the author used. Which is why Cork is so meticulous 317 00:14:45,570 --> 00:14:48,470 in his robotics toolbox. When you define a serial 318 00:14:48,470 --> 00:14:50,789 link object in Python, you have to explicitly 319 00:14:50,789 --> 00:14:53,169 pass the kinematic parameters and declare whether 320 00:14:53,169 --> 00:14:56,100 they are standard or modified. You can't just 321 00:14:56,100 --> 00:14:59,429 leave it to chance. Never. But this highlights 322 00:14:59,429 --> 00:15:01,690 the fundamental nature of forward kinematics. 323 00:15:02,309 --> 00:15:05,929 It is a forward function. Meaning? You input 324 00:15:05,929 --> 00:15:08,830 joint angles, you output a unique pose. It's 325 00:15:08,830 --> 00:15:12,460 a one -way street. But in the real world of manufacturing, 326 00:15:13,019 --> 00:15:15,519 we almost never have that luxury. We are faced 327 00:15:15,519 --> 00:15:18,720 with the exact inverse problem. OK, let's unpack 328 00:15:18,720 --> 00:15:21,379 this. OK. Because if I am programming a robot 329 00:15:21,379 --> 00:15:25,230 to spot weld a car chassis, I don't care what 330 00:15:25,230 --> 00:15:27,730 the angle of joint three is, right? Not at all. 331 00:15:27,950 --> 00:15:30,649 I just have an XYZ coordinate on the chassis 332 00:15:30,649 --> 00:15:33,029 and a specific orientation the welding torch 333 00:15:33,029 --> 00:15:36,029 needs to hold. I know the destination. I need 334 00:15:36,029 --> 00:15:38,110 the software to tell me what joint angles will 335 00:15:38,110 --> 00:15:41,190 get me there. This is inverse kinematics, IK. 336 00:15:41,649 --> 00:15:44,509 And inverse kinematics is a vastly more difficult 337 00:15:44,509 --> 00:15:46,690 mathematical problem. It's like giving someone 338 00:15:46,690 --> 00:15:49,090 driving directions step by step versus just giving 339 00:15:49,090 --> 00:15:51,389 them a destination address and forcing them to 340 00:15:51,389 --> 00:15:53,740 calculate the route backwards. That's a great 341 00:15:53,740 --> 00:15:55,980 way to think about it. Forward kinematics is 342 00:15:55,980 --> 00:15:58,720 a simple algebraic substitution. It is a one 343 00:15:58,720 --> 00:16:02,379 -to -one mapping. Inverse kinematics is heavily 344 00:16:02,379 --> 00:16:04,960 nonlinear, and crucially, it is not a unique 345 00:16:04,960 --> 00:16:07,580 mapping. Wait, hold on. Why isn't it unique? 346 00:16:07,700 --> 00:16:10,559 If I want the torch at a specific point in space, 347 00:16:10,899 --> 00:16:13,399 shouldn't there be a specific set of joint angles 348 00:16:13,399 --> 00:16:15,960 that puts it there? Think about your own body. 349 00:16:16,879 --> 00:16:19,740 Stand facing a wall and place your palm flat 350 00:16:19,740 --> 00:16:22,830 against a specific spot on the wall. Okay, palm 351 00:16:22,830 --> 00:16:25,149 on the wall. Keep your hand completely still. 352 00:16:25,190 --> 00:16:28,269 That is your constrained end effector pose. Now, 353 00:16:28,490 --> 00:16:30,789 without moving your hand, you can raise your 354 00:16:30,789 --> 00:16:33,610 elbow high up into the air, or you can drop your 355 00:16:33,610 --> 00:16:38,659 elbow down close to your ribs. Oh. The hand's 356 00:16:38,659 --> 00:16:40,720 pose didn't change, but the joint angles in my 357 00:16:40,720 --> 00:16:43,519 shoulder and elbow completely changed. Precisely. 358 00:16:44,610 --> 00:16:46,750 A standard six -axis robot arm usually has up 359 00:16:46,750 --> 00:16:48,809 to eight different mathematical solutions to 360 00:16:48,809 --> 00:16:51,789 reach a single pose in its workspace. Eight? 361 00:16:52,149 --> 00:16:55,049 Yes. In industry, these are categorized into 362 00:16:55,049 --> 00:16:58,289 configuration stays. You have elbow up versus 363 00:16:58,289 --> 00:17:02,049 elbow down, lefty versus righty, and wrist flipped 364 00:17:02,049 --> 00:17:05,690 versus unflipped. So the IDK solver doesn't just 365 00:17:05,690 --> 00:17:09,170 hand you an answer, it hands you a menu of eight 366 00:17:09,170 --> 00:17:11,710 answers, and your software has to decide which 367 00:17:11,710 --> 00:17:14,970 one to use. And that decision is absolutely critical. 368 00:17:15,150 --> 00:17:17,849 Why? Because if your robot is currently in an 369 00:17:17,849 --> 00:17:20,660 elbow up configuration, And the next point on 370 00:17:20,660 --> 00:17:22,859 your welding path is calculated using an elbow 371 00:17:22,859 --> 00:17:26,480 down solution. Oh no. The robot is going to wildly 372 00:17:26,480 --> 00:17:29,079 swing its entire arm through the workspace just 373 00:17:29,079 --> 00:17:31,619 to move the torch one millimeter. It has to completely 374 00:17:31,619 --> 00:17:33,700 reconfigure its posture to switch solutions. 375 00:17:33,880 --> 00:17:35,940 That would destroy whatever it's working on. 376 00:17:36,180 --> 00:17:37,940 Exactly. It would rip the welding torch right 377 00:17:37,940 --> 00:17:41,839 off. So the IK algorithm must track the configuration 378 00:17:41,839 --> 00:17:44,329 state. and choose the solution that minimizes 379 00:17:44,329 --> 00:17:47,329 joint travel from the current pose while also 380 00:17:47,329 --> 00:17:49,509 checking for joint limits and self -collision. 381 00:17:49,950 --> 00:17:52,309 It's a lot of plates to keep spinning. So Cork 382 00:17:52,309 --> 00:17:54,769 breaks down two primary ways to actually solve 383 00:17:54,769 --> 00:17:58,190 these IK equations, right? Closed form analytic 384 00:17:58,190 --> 00:18:01,109 solutions and iterative numerical solutions. 385 00:18:01,670 --> 00:18:03,309 Yes. Closed form sounds like the Holy Grail. 386 00:18:03,349 --> 00:18:05,730 It means you actually solve the algebra on paper, 387 00:18:05,910 --> 00:18:08,509 right? Yes. A closed form solution means you 388 00:18:08,509 --> 00:18:11,559 have derived a set of explicit deterministic 389 00:18:11,559 --> 00:18:13,980 algebraic equations. No guessing. No guessing. 390 00:18:14,930 --> 00:18:17,230 You input the desired four by four pose matrix 391 00:18:17,230 --> 00:18:20,049 and the equation spit out the exact joint angles 392 00:18:20,049 --> 00:18:23,470 in microseconds. It is computationally incredibly 393 00:18:23,470 --> 00:18:26,049 efficient. But I hear a but coming. There is 394 00:18:26,049 --> 00:18:28,589 a massive catch. The geometry has to support 395 00:18:28,589 --> 00:18:32,220 it. The geometry has to be highly specific. In 396 00:18:32,220 --> 00:18:34,980 1968, a mathematician named Pieper Prove that 397 00:18:34,980 --> 00:18:37,319 for a six degree of freedom serial manipulator, 398 00:18:37,640 --> 00:18:40,039 you can only find a closed form analytic solution 399 00:18:40,039 --> 00:18:43,019 if three consecutive revolute joints intersect 400 00:18:43,019 --> 00:18:45,279 at a single point in space. A spherical wrist. 401 00:18:45,660 --> 00:18:48,119 Exactly. Joints four, five, and six must all 402 00:18:48,119 --> 00:18:50,079 physically intersect. Why is that intersection 403 00:18:50,079 --> 00:18:53,180 so magic? This geometric trick mathematically 404 00:18:53,180 --> 00:18:56,079 decouples the position problem from the orientation 405 00:18:56,079 --> 00:18:58,799 problem. Oh, it splits the math in half. Yes. 406 00:18:59,140 --> 00:19:01,880 You use the first three joints, the arm. to place 407 00:19:01,880 --> 00:19:03,759 the wrist's intersection point at the correct 408 00:19:03,759 --> 00:19:07,420 X, Y, Z coordinate. Then you use the last three 409 00:19:07,420 --> 00:19:10,099 joints, the spherical wrist, purely to pivot 410 00:19:10,099 --> 00:19:12,059 around that point and achieve the correct orientation. 411 00:19:12,460 --> 00:19:15,059 It breaks a crazy PsyXD problem into two manageable 412 00:19:15,059 --> 00:19:17,740 3D problems. It does. And this is why almost 413 00:19:17,740 --> 00:19:20,000 every major industrial 6 -axis robot for the 414 00:19:20,000 --> 00:19:23,680 last 50 years, the Puma, the KUKA KR series, 415 00:19:23,900 --> 00:19:26,700 the FANUC models, are explicitly designed with 416 00:19:26,700 --> 00:19:28,980 spherical wrists. The hardware was literally 417 00:19:28,980 --> 00:19:31,539 designed to accommodate the limitations of 20th 418 00:19:31,539 --> 00:19:34,160 century microprocessors. Exactly. They needed 419 00:19:34,160 --> 00:19:36,819 the math to be closed form so it could run fast 420 00:19:36,819 --> 00:19:39,720 enough on primitive computers. But what if I'm 421 00:19:39,720 --> 00:19:42,259 not using a factory six -axis arm? What if I'm 422 00:19:42,259 --> 00:19:44,240 using, you know, a modern collaborative robot 423 00:19:44,240 --> 00:19:47,180 that has offset joints or a seven -axis arm that 424 00:19:47,180 --> 00:19:49,299 has an extra degree of freedom? Peeper's condition 425 00:19:49,299 --> 00:19:52,500 fails then, right? It fails completely. The algebra 426 00:19:52,500 --> 00:19:55,220 becomes unsolvable in a closed form. So what 427 00:19:55,220 --> 00:19:58,369 do you do? When you lose the closed form solution, 428 00:19:58,750 --> 00:20:01,049 you have to rely on numerical inverse kinematics. 429 00:20:01,529 --> 00:20:03,910 You treat it as an optimization problem. This 430 00:20:03,910 --> 00:20:06,089 is essentially gradient descent for robotics, 431 00:20:06,210 --> 00:20:08,470 right? You take a guess, measure the error, and 432 00:20:08,470 --> 00:20:10,869 tweak the joints. That is exactly what it is. 433 00:20:11,210 --> 00:20:12,950 You start with an initial guess for the joint 434 00:20:12,950 --> 00:20:15,410 angles, usually wherever the robot is currently 435 00:20:15,410 --> 00:20:18,930 positioned. Makes sense. You run forward kinematics 436 00:20:18,930 --> 00:20:20,930 to see where the end effector ends up with that 437 00:20:20,930 --> 00:20:23,549 guess. Then you calculate the error vector between 438 00:20:23,549 --> 00:20:26,650 that pose and your actual target pose. Like I'm 439 00:20:26,650 --> 00:20:29,549 two inches too far to the left. Right. Then use 440 00:20:29,549 --> 00:20:32,430 numerical methods like Newton -Rofson or Gauss 441 00:20:32,430 --> 00:20:35,509 -Newton or Levenberg -Marquardt, which Cork implements 442 00:20:35,509 --> 00:20:37,730 in his toolbox to calculate a small adjustment 443 00:20:37,730 --> 00:20:40,549 to the joint angles that will step the end defector 444 00:20:40,549 --> 00:20:43,029 closer to the target. And you just loop that 445 00:20:43,029 --> 00:20:45,890 over and over until the error is within an acceptable 446 00:20:45,890 --> 00:20:48,970 tolerance. Exactly. But wait, that requires computing 447 00:20:48,970 --> 00:20:51,109 the gradient. You have to know which way tweaking 448 00:20:51,109 --> 00:20:53,430 a joint will actually move the hand. Yes, you 449 00:20:53,430 --> 00:20:56,079 do. And that perfectly transitions us from static 450 00:20:56,079 --> 00:20:59,240 geometry to the calculus of motion. Because to 451 00:20:59,240 --> 00:21:01,440 compute that gradient, to understand how a change 452 00:21:01,440 --> 00:21:03,960 in joint angle produces a change in spatial pose, 453 00:21:04,339 --> 00:21:06,299 we have to look at velocities. We have to open 454 00:21:06,299 --> 00:21:09,359 chapter 8. Manipulator velocity. Let's do it. 455 00:21:10,140 --> 00:21:12,380 This is the most profound mathematical leap in 456 00:21:12,380 --> 00:21:15,880 the textbook. In chapter 7, we lived in the configuration 457 00:21:15,880 --> 00:21:19,740 space. Theta maps to pose. Now we are entering 458 00:21:19,740 --> 00:21:22,319 differential kinematics. The rate of change. 459 00:21:22,500 --> 00:21:25,059 Yes, the rate of change of theta maps to the 460 00:21:25,059 --> 00:21:27,839 spatial velocity of the pose. So instead of asking 461 00:21:27,839 --> 00:21:30,940 where is the hand, we are asking if I spin the 462 00:21:30,940 --> 00:21:33,880 shoulder motor at five radians per second, how 463 00:21:33,880 --> 00:21:35,859 fast is the hand moving through space and in 464 00:21:35,859 --> 00:21:38,700 what direction? And the mathematical entity that 465 00:21:38,700 --> 00:21:41,200 translates joint velocity into spatial velocity 466 00:21:41,200 --> 00:21:45,000 is the manipulator Jacobian. The Jacobian. We 467 00:21:45,000 --> 00:21:46,859 briefly encountered the concept of a Jacobian 468 00:21:46,859 --> 00:21:49,519 in part one when discussing simple motion mapping, 469 00:21:49,519 --> 00:21:53,710 but here Quark applies it fully to complex manipulator 470 00:21:53,710 --> 00:21:56,190 chains. Let's really drill down into what this 471 00:21:56,190 --> 00:21:58,430 matrix is physically representing, because it's 472 00:21:58,430 --> 00:22:01,829 so important. The Jacobian is a matrix of partial 473 00:22:01,829 --> 00:22:04,589 derivatives. For a six -joint robot operating 474 00:22:04,589 --> 00:22:07,809 in a six degree of freedom task space, the Jacobian 475 00:22:07,809 --> 00:22:10,890 is a six by six matrix. OK, six rows, six columns. 476 00:22:11,509 --> 00:22:14,630 What goes in and what comes out? The input is 477 00:22:14,630 --> 00:22:17,460 a six by one vector of joint velocities. how 478 00:22:17,460 --> 00:22:20,299 fast each motor is spinning. The output is a 479 00:22:20,299 --> 00:22:23,519 6x1 vector representing spatial velocity. The 480 00:22:23,519 --> 00:22:26,660 linear velocity is along x, y, and z, and the 481 00:22:26,660 --> 00:22:29,259 angular velocity is around x, y, and z. So if 482 00:22:29,259 --> 00:22:32,039 I multiply the joint velocity vector by the Jacobian 483 00:22:32,039 --> 00:22:35,200 matrix, it instantly translates motor speed into 484 00:22:35,200 --> 00:22:38,980 tool speed. Yes. But here's the critical insight. 485 00:22:39,539 --> 00:22:43,099 The Jacobian is not a constant. It changes? It 486 00:22:43,099 --> 00:22:45,279 is entirely dependent on the robot's current 487 00:22:45,279 --> 00:22:48,440 pose. Right, because the leverage changes. If 488 00:22:48,440 --> 00:22:51,019 my arm is bent at 90 degrees, spinning my shoulder 489 00:22:51,019 --> 00:22:53,680 motor moves my hand forward. But if my arm is 490 00:22:53,680 --> 00:22:55,180 pointing straight up at the ceiling, spinning 491 00:22:55,180 --> 00:22:57,500 that exact same shoulder motor at the exact same 492 00:22:57,500 --> 00:23:00,940 speed now sweeps my hand sideways. Exactly. The 493 00:23:00,940 --> 00:23:03,640 mapping from motor to space depends entirely 494 00:23:03,640 --> 00:23:06,319 on the geometry at that specific millisecond. 495 00:23:06,599 --> 00:23:09,359 You have to recompute the Jacobian matrix continuously 496 00:23:09,359 --> 00:23:12,559 as the robot moves through its trajectory. That 497 00:23:12,559 --> 00:23:15,349 sounds like a lot of math. It is, but once you 498 00:23:15,349 --> 00:23:18,009 have it, it unlocks a control paradigm called 499 00:23:18,009 --> 00:23:22,849 Resolved Rate Motion Control, or RRMC. This is 500 00:23:22,849 --> 00:23:25,589 where we bypass the headache of inverse kinematics 501 00:23:25,589 --> 00:23:28,150 entirely for path following, isn't it? It is. 502 00:23:28,490 --> 00:23:30,710 If you want a robot to draw a perfectly straight 503 00:23:30,710 --> 00:23:34,450 line with a marker, doing numerical IK for thousands 504 00:23:34,450 --> 00:23:37,769 of microscopic points along that line is computationally 505 00:23:37,769 --> 00:23:40,470 exhausting. And didn't you say it can lead to 506 00:23:40,470 --> 00:23:43,329 jerky motion? Yes. If the solver jumps between 507 00:23:43,329 --> 00:23:45,269 different configuration states, like suddenly 508 00:23:45,269 --> 00:23:48,150 flipping from elbow up to elbow down, the motion 509 00:23:48,150 --> 00:23:51,970 is ruined. So how does RRMC solve that? Instead 510 00:23:51,970 --> 00:23:53,970 of commanding position, you command velocity. 511 00:23:54,829 --> 00:23:56,450 You know you want the marker to move purely in 512 00:23:56,450 --> 00:23:59,769 the positive x direction at 0 .1 meters per second. 513 00:24:00,569 --> 00:24:03,109 Your desired spatial velocity vector is simply 514 00:24:03,109 --> 00:24:10,000 0 .1000000. Just pure X velocity. To find out 515 00:24:10,000 --> 00:24:11,980 how fast to spin the motors to achieve that, 516 00:24:13,000 --> 00:24:15,440 you take the inverse of the Jacobian matrix and 517 00:24:15,440 --> 00:24:17,619 you multiply it by your desired spatial velocity 518 00:24:17,619 --> 00:24:22,000 vector. That is so elegant. The inverse Jacobian 519 00:24:22,000 --> 00:24:24,779 acts as a perfect translator in reverse. You 520 00:24:24,779 --> 00:24:27,099 just feed it the Cartesian speed you want and 521 00:24:27,099 --> 00:24:30,119 it outputs the exact motor speeds required to 522 00:24:30,119 --> 00:24:33,339 track that line flawlessly. It is beautiful linear 523 00:24:33,339 --> 00:24:35,869 algebra. You integrate those joint velocities 524 00:24:35,869 --> 00:24:38,710 over time to actually move the motors, and the 525 00:24:38,710 --> 00:24:41,029 end effector glides perfectly along the path. 526 00:24:41,710 --> 00:24:44,210 But anyone who remembers their linear algebra 527 00:24:44,210 --> 00:24:46,730 class is screaming at the player right now. Oh, 528 00:24:46,730 --> 00:24:49,289 I know what you're gonna say. You cannot always 529 00:24:49,289 --> 00:24:52,230 invert a matrix. No, you cannot. A matrix has 530 00:24:52,230 --> 00:24:54,890 to be full rank. Its determinant cannot be zero. 531 00:24:55,470 --> 00:24:57,589 If the determinant of the Jacobian hits zero, 532 00:24:57,779 --> 00:25:01,039 the inverse explodes to infinity. And that mathematical 533 00:25:01,039 --> 00:25:04,140 explosion is the definition of a kinematic singularity, 534 00:25:04,700 --> 00:25:06,680 which brings us right back to the disaster scenario 535 00:25:06,680 --> 00:25:09,440 you painted at the start of this deep dive. A 536 00:25:09,440 --> 00:25:13,700 geometric black hole. So what actually happens 537 00:25:13,700 --> 00:25:17,720 to the physical robot at a singularity? A singularity 538 00:25:17,720 --> 00:25:20,319 occurs when the robot's links align in such a 539 00:25:20,319 --> 00:25:22,559 way that the columns of the Jacobian matrix are 540 00:25:22,559 --> 00:25:26,930 no longer linearly independent. Meaning? It means 541 00:25:26,930 --> 00:25:29,069 the robot has physically lost one or more degrees 542 00:25:29,069 --> 00:25:31,869 of freedom in the task space. Here's where it 543 00:25:31,869 --> 00:25:34,690 gets really interesting. Let's make this tangible. 544 00:25:35,009 --> 00:25:37,690 I want you to imagine holding your arm perfectly 545 00:25:37,690 --> 00:25:39,869 straight out in front of you, locking your elbow, 546 00:25:40,309 --> 00:25:42,430 and trying to reach an object that is one inch 547 00:25:42,430 --> 00:25:44,630 further away directly in front of your fingertips. 548 00:25:45,029 --> 00:25:48,259 That is a classic boundary singularity. Your 549 00:25:48,259 --> 00:25:50,480 shoulder joint works perfectly fine. You can 550 00:25:50,480 --> 00:25:52,839 swing your straight arm up, down, left, right, 551 00:25:53,039 --> 00:25:55,519 but because your elbow is locked straight, no 552 00:25:55,519 --> 00:25:57,839 combination of those shoulder movements will 553 00:25:57,839 --> 00:26:00,339 move your fingertips straightforward. You have 554 00:26:00,339 --> 00:26:03,539 absolutely no mechanical leverage in that specific 555 00:26:03,539 --> 00:26:07,500 Cartesian direction. Now, apply the math of resolved 556 00:26:07,500 --> 00:26:09,900 rate motion control to that physical reality. 557 00:26:10,420 --> 00:26:13,259 Okay, so if your software demands a spatial velocity 558 00:26:13,259 --> 00:26:17,710 of 0 .1 comma 0 comma 0 straight forward, The 559 00:26:17,710 --> 00:26:20,029 inverse Jacobian tries to calculate the required 560 00:26:20,029 --> 00:26:22,950 joint speeds. But because you have zero leverage, 561 00:26:23,309 --> 00:26:25,289 the math dictates that the elbow or shoulder 562 00:26:25,289 --> 00:26:28,410 motors must spin at infinite velocity to achieve 563 00:26:28,410 --> 00:26:31,269 even a millimeter of forward motion. And since 564 00:26:31,269 --> 00:26:33,690 motors can't spin at infinity, the calculated 565 00:26:33,690 --> 00:26:36,569 current spikes, the amplifier saturates, the 566 00:26:36,569 --> 00:26:39,730 software faults, and the robot physically violently 567 00:26:39,730 --> 00:26:43,349 shutters or shuts down. Exactly. But Cork is 568 00:26:43,349 --> 00:26:45,529 very careful to distinguish between the loss 569 00:26:45,529 --> 00:26:48,750 of instantaneous task motion directions and the 570 00:26:48,750 --> 00:26:51,529 physical locking of joints. This is a huge misconception. 571 00:26:52,130 --> 00:26:54,170 Let's clarify that. The joints don't actually 572 00:26:54,170 --> 00:26:57,089 freeze, right? Right. At a singularity, the motors 573 00:26:57,089 --> 00:27:00,069 are perfectly fine. They can spin freely. The 574 00:27:00,069 --> 00:27:02,630 mechanical bearings aren't seized. So what's 575 00:27:02,630 --> 00:27:05,009 the problem? The robot simply cannot combine 576 00:27:05,009 --> 00:27:07,970 the motion of those freely spinning joints to 577 00:27:07,970 --> 00:27:11,309 produce a velocity vector in one specific, impossible 578 00:27:11,309 --> 00:27:13,859 direction. If you command the robot to move sideways 579 00:27:13,859 --> 00:27:16,500 instead, it will move perfectly. The limitation 580 00:27:16,500 --> 00:27:19,680 is entirely directional. So as an engineer, you 581 00:27:19,680 --> 00:27:21,940 have to design trajectories that keep the robot 582 00:27:21,940 --> 00:27:24,119 far away from these singular configurations. 583 00:27:24,640 --> 00:27:26,799 You need a metric to measure how close you are 584 00:27:26,799 --> 00:27:29,180 to the edge of the cliff. And Cork introduces 585 00:27:29,180 --> 00:27:31,819 Yoshikawa's manipulability measure for this. 586 00:27:32,200 --> 00:27:34,960 Ah, yes. Yoshikawa proposed a way to visualize 587 00:27:34,960 --> 00:27:37,700 the dexterity of a robot at any given pose using 588 00:27:37,700 --> 00:27:40,200 something called the velocity ellipsoid. The 589 00:27:40,200 --> 00:27:43,359 velocity ellipsoid. Yeah. Let's unpack the geometry 590 00:27:43,359 --> 00:27:47,079 of that. Imagine a unit hypersphere in the joint 591 00:27:47,079 --> 00:27:50,940 velocity space. A perfect sphere. Yes. This represents 592 00:27:50,940 --> 00:27:54,240 all the motors spinning at some normalized maximum 593 00:27:54,240 --> 00:27:57,460 speed in every possible combination. The Jacobian 594 00:27:57,460 --> 00:28:00,299 matrix maps that perfect sphere into the Cartesian 595 00:28:00,299 --> 00:28:02,799 task space. But because the robot has different 596 00:28:02,799 --> 00:28:04,839 leverage and different directions, it doesn't 597 00:28:04,839 --> 00:28:07,579 stay a sphere, does it? No. It gets distorted 598 00:28:07,579 --> 00:28:11,410 into a multi -dimensional ellipsoid. Like an 599 00:28:11,410 --> 00:28:14,490 egg shape? Exactly. The shape of that ellipsoid 600 00:28:14,490 --> 00:28:16,470 tells you everything about the robot's capabilities 601 00:28:16,470 --> 00:28:20,109 in that specific pose. The longest axis of the 602 00:28:20,109 --> 00:28:22,630 ellipsoid points in the direction where the robot 603 00:28:22,630 --> 00:28:25,710 can move the fastest, where it has the most mechanical 604 00:28:25,710 --> 00:28:28,509 advantage. And the short axis? The shortest axis 605 00:28:28,509 --> 00:28:31,029 points in the direction where the robot struggles 606 00:28:31,029 --> 00:28:34,349 to move. So going back to our straight arm analogy, 607 00:28:34,890 --> 00:28:37,069 if your arm is locked perfectly straight out, 608 00:28:37,140 --> 00:28:40,039 That velocity ellipsoid gets crushed completely 609 00:28:40,039 --> 00:28:42,619 flat in the forward direction. It becomes a 2D 610 00:28:42,619 --> 00:28:45,000 pancake. Its volume shrinks to absolute zero. 611 00:28:45,259 --> 00:28:48,500 And that is Yoshikawa's measure. It is proportional 612 00:28:48,500 --> 00:28:51,099 to the volume of the velocity ellipsoid. How 613 00:28:51,099 --> 00:28:53,700 do you calculate that? Mathematically, it's calculated 614 00:28:53,700 --> 00:28:55,920 using the square root of the determinant of the 615 00:28:55,920 --> 00:28:59,460 Jacobian multiplied by its transpose. A large 616 00:28:59,460 --> 00:29:02,319 volume means high manipulability. The ellipsoid 617 00:29:02,319 --> 00:29:05,480 is plump and round, and the robot can move dexterously 618 00:29:05,480 --> 00:29:08,369 in any direction. But a zero volume means the 619 00:29:08,369 --> 00:29:10,970 ellipsoid has collapsed and you are in a singularity. 620 00:29:11,470 --> 00:29:14,549 Exactly. So your path planter is constantly monitoring 621 00:29:14,549 --> 00:29:16,950 Yoshikawa's measure, making sure the volume doesn't 622 00:29:16,950 --> 00:29:19,509 drop too low. Now, we established earlier that 623 00:29:19,509 --> 00:29:22,250 it takes six joints to fully command a six degree 624 00:29:22,250 --> 00:29:25,890 of freedom task space, but cork dashgates significant 625 00:29:25,890 --> 00:29:28,869 time to over actuation, redundant robots. Just 626 00:29:28,869 --> 00:29:31,980 like the ABB Yumi. Right, which has two 7 -axis 627 00:29:31,980 --> 00:29:35,079 arms, and the OC robotic snake arm, which has 628 00:29:35,079 --> 00:29:37,500 20 degrees of freedom. Why on earth would you 629 00:29:37,500 --> 00:29:40,680 add the weight, cost, and complexity of 20 joints 630 00:29:40,680 --> 00:29:42,779 when the math says you only need 6? Because when 631 00:29:42,779 --> 00:29:44,920 you have more joints than task space dimensions, 632 00:29:45,480 --> 00:29:48,200 your Jacobian matrix is no longer square. It's 633 00:29:48,200 --> 00:29:53,400 a 6 by n matrix, where n is 7 or 20. And in linear 634 00:29:53,400 --> 00:29:55,799 algebra? An under -determined system like that 635 00:29:55,799 --> 00:29:59,029 possesses a null space. The null space. This 636 00:29:59,029 --> 00:30:01,369 is where redundant robotics feels like magic. 637 00:30:01,730 --> 00:30:04,170 What does projecting motion into the null space 638 00:30:04,170 --> 00:30:06,609 physically look like? It means there is an infinite 639 00:30:06,609 --> 00:30:08,690 set of joint velocity combinations that result 640 00:30:08,690 --> 00:30:12,390 in exactly zero spatial velocity at the end effector. 641 00:30:12,569 --> 00:30:15,190 Wait, so the motors are spinning, the joints 642 00:30:15,190 --> 00:30:17,690 are moving, but the tool tip is completely frozen 643 00:30:17,690 --> 00:30:20,349 in space. Yes. Imagine you are holding a full 644 00:30:20,349 --> 00:30:22,890 cup of coffee. Your hand is the end effector. 645 00:30:23,069 --> 00:30:25,329 constrained to keep the cup level and stationary. 646 00:30:25,809 --> 00:30:28,009 Because your human arm has seven degrees of freedom, 647 00:30:28,529 --> 00:30:30,849 you can raise your elbow up, swing it to the 648 00:30:30,849 --> 00:30:34,049 side, and drop it down, all without moving the 649 00:30:34,049 --> 00:30:37,069 coffee cup a single millimeter. You are executing 650 00:30:37,069 --> 00:30:39,869 internal reconfigurations in the null space of 651 00:30:39,869 --> 00:30:43,130 your arm's Jacobian. Exactly. For an industrial 652 00:30:43,130 --> 00:30:45,250 robot, that means it can hold a welding torch 653 00:30:45,250 --> 00:30:47,890 perfectly steady on a seam while simultaneously 654 00:30:47,890 --> 00:30:50,170 sweeping its elbow up to dodge a moving conveyor 655 00:30:50,170 --> 00:30:53,019 belt. or reconfiguring its joints to maximize 656 00:30:53,019 --> 00:30:55,440 Yoshikawa's manipulability and stay away from 657 00:30:55,440 --> 00:30:58,039 a singularity. It is the ultimate flexibility. 658 00:30:58,740 --> 00:31:01,500 For a 20 -joint snake arm inspecting the inside 659 00:31:01,500 --> 00:31:04,180 of a nuclear reactor, it can snake its entire 660 00:31:04,180 --> 00:31:07,420 body through a maze of pipes, constantly reconfiguring 661 00:31:07,420 --> 00:31:09,960 its 20 joints in the null space, while keeping 662 00:31:09,960 --> 00:31:12,299 the inspection camera on the tip perfectly stable 663 00:31:12,299 --> 00:31:15,319 and tracking a linear path. The math allows you 664 00:31:15,319 --> 00:31:17,400 to separate the primary task, what the tool is 665 00:31:17,400 --> 00:31:19,900 doing, from the secondary task, what the body 666 00:31:19,900 --> 00:31:22,369 is doing. It's brilliant. It's beautiful math. 667 00:31:22,769 --> 00:31:26,890 But now there is one final mind -bending property 668 00:31:26,890 --> 00:31:29,329 of the Jacobian that Cork covers in Chapter 8, 669 00:31:29,349 --> 00:31:31,529 and I have to push back on this because geometrically 670 00:31:31,529 --> 00:31:35,569 it feels wrong. Uh -oh. What's the problem? We 671 00:31:35,569 --> 00:31:37,730 have established that the Jacobian maps velocities. 672 00:31:37,990 --> 00:31:40,509 It is a matrix of derivatives describing motion. 673 00:31:41,150 --> 00:31:43,630 But Cork says that if you take the transpose 674 00:31:43,630 --> 00:31:46,710 of the Jacobian, it suddenly maps static forces. 675 00:31:47,150 --> 00:31:50,069 How on earth does flipping a speed matrix give 676 00:31:50,069 --> 00:31:53,650 you static forces? It seems highly counterintuitive 677 00:31:53,650 --> 00:31:55,509 until you look at the conservation of energy, 678 00:31:55,789 --> 00:31:58,549 specifically the principle of virtual work. Okay, 679 00:31:58,549 --> 00:32:00,769 walk me through the physics of that. The total 680 00:32:00,769 --> 00:32:03,769 work done by the robot's joints must equal the 681 00:32:03,769 --> 00:32:06,190 total work done by the end effector on the environment. 682 00:32:06,930 --> 00:32:09,529 Work is force multiplied by distance, right? 683 00:32:09,789 --> 00:32:13,190 Yes. In the joint space, work is torque, which 684 00:32:13,190 --> 00:32:16,009 we call tau, multiplied by angular displacement, 685 00:32:16,190 --> 00:32:19,420 d theta. In the task space, work is the force 686 00:32:19,420 --> 00:32:22,339 vector F multiplied by the spatial displacement 687 00:32:22,339 --> 00:32:28,240 dx. So tau multiplied by d theta equals F multiplied 688 00:32:28,240 --> 00:32:32,140 by dx. Exactly. Now, remember our differential 689 00:32:32,140 --> 00:32:34,900 kinematics equation. The spatial displacement 690 00:32:34,900 --> 00:32:38,380 dx is equal to the Jacobian matrix J multiplied 691 00:32:38,380 --> 00:32:41,410 by the joint displacement d theta. Okay, I follow. 692 00:32:41,829 --> 00:32:43,769 If you substitute that into the work equation 693 00:32:43,769 --> 00:32:46,990 and cancel out the d theta on both sides, Oh, 694 00:32:47,630 --> 00:32:49,730 the linear algebra shakes out and you are left 695 00:32:49,730 --> 00:32:53,269 with torque equals the Jacobian transpose multiplied 696 00:32:53,269 --> 00:32:58,089 by force. Tau equals JT times F. That is it. 697 00:32:58,549 --> 00:33:00,710 The Jacobian transpose is the mathematical bridge 698 00:33:00,710 --> 00:33:04,329 between task space forces and joint space torques. 699 00:33:04,609 --> 00:33:07,970 So if a human pushes on the robot's end effector 700 00:33:07,970 --> 00:33:11,230 with a 50 Newton force vector. You multiply that 701 00:33:11,230 --> 00:33:14,630 force vector by the Jacobian transpose. And it 702 00:33:14,630 --> 00:33:17,250 instantly calculates exactly how many newton 703 00:33:17,250 --> 00:33:19,970 meters of torque every single motor will experience 704 00:33:19,970 --> 00:33:23,210 as a result of that push. That is incredibly 705 00:33:23,210 --> 00:33:26,009 powerful. It means you can measure the current 706 00:33:26,009 --> 00:33:28,190 spikes on your joint motors, pass them through 707 00:33:28,190 --> 00:33:30,849 the inverse of the Jacobian transpose, and calculate 708 00:33:30,849 --> 00:33:33,509 exactly how hard the tooltip is pressing against 709 00:33:33,509 --> 00:33:35,869 a table without even needing a force sensor on 710 00:33:35,869 --> 00:33:39,440 the wrist. Yes. And Cork shows how this transpose 711 00:33:39,440 --> 00:33:41,980 property gives us the most elegant method for 712 00:33:41,980 --> 00:33:45,220 solving numerical inverse kinematics. We talked 713 00:33:45,220 --> 00:33:47,599 about iterative tweaking earlier, right? And 714 00:33:47,599 --> 00:33:49,960 this raises an important question. How do we 715 00:33:49,960 --> 00:33:52,819 actually do that tweaking? The Jacobian transpose 716 00:33:52,819 --> 00:33:56,099 provides a physical interpretation for it, the 717 00:33:56,099 --> 00:33:59,220 virtual spring. How does a virtual spring solve 718 00:33:59,220 --> 00:34:02,599 IK? Imagine the robot's end effector is at its 719 00:34:02,599 --> 00:34:05,640 current position, but you want it to be 10 centimeters 720 00:34:05,640 --> 00:34:09,130 away. You mathematically simulate a physical 721 00:34:09,130 --> 00:34:12,010 spring stretched between the current hand position 722 00:34:12,010 --> 00:34:15,250 and the target position. Okay, so this virtual 723 00:34:15,250 --> 00:34:18,469 spring is pulling on the hand, generating a virtual 724 00:34:18,469 --> 00:34:21,269 force vector pointing toward the target. We take 725 00:34:21,269 --> 00:34:24,050 that virtual force vector and we multiply it 726 00:34:24,050 --> 00:34:27,239 by the Jacobian transpose. That gives us a set 727 00:34:27,239 --> 00:34:30,579 of virtual joint torques. We then simulate the 728 00:34:30,579 --> 00:34:33,260 joints moving in the direction of those torques. 729 00:34:33,440 --> 00:34:35,699 And as they move, the hand gets closer to the 730 00:34:35,699 --> 00:34:38,909 target. The virtual spring relaxes, the virtual 731 00:34:38,909 --> 00:34:41,829 force decreases, and the joint speeds taper off. 732 00:34:42,389 --> 00:34:45,349 We run the simulation in a tight loop. When the 733 00:34:45,349 --> 00:34:48,110 spring force hits zero, the hand is at the target, 734 00:34:48,530 --> 00:34:51,010 and we just read the final joint angles. I mean, 735 00:34:51,070 --> 00:34:53,889 it's clever, but I have to challenge the practicality 736 00:34:53,889 --> 00:34:56,170 of this. You are constantly running a physics 737 00:34:56,170 --> 00:34:59,170 simulation of a fake spring just to figure out 738 00:34:59,170 --> 00:35:02,409 where your hand is. Isn't that massively computationally 739 00:35:02,409 --> 00:35:04,769 expensive compared to a closed -form algebraic 740 00:35:04,769 --> 00:35:08,769 equation? is more expensive. A closed -form solution 741 00:35:08,769 --> 00:35:12,110 takes a few microseconds. A numerical Jacobian 742 00:35:12,110 --> 00:35:14,849 transpose loop might take a millisecond. In the 743 00:35:14,849 --> 00:35:17,849 1980s, that was a deal -breaker. But today? Today, 744 00:35:17,849 --> 00:35:20,730 with modern processors, that millisecond is negligible. 745 00:35:21,389 --> 00:35:24,550 And the robustness it provides, the ability to 746 00:35:24,550 --> 00:35:27,510 easily handle seven -axis redundant robots, to 747 00:35:27,510 --> 00:35:30,469 easily avoid joint limits by adding virtual repulsive 748 00:35:30,469 --> 00:35:32,869 springs to the joints, makes it the standard 749 00:35:32,869 --> 00:35:35,460 approach in modern robotics frameworks. But notice 750 00:35:35,460 --> 00:35:37,719 the assumption we made in that simulation. We 751 00:35:37,719 --> 00:35:40,340 assume that when we applied the virtual torques, 752 00:35:40,559 --> 00:35:43,219 the joints just effortlessly moved. We assume 753 00:35:43,219 --> 00:35:46,360 the robot was made of massless lines. Right. 754 00:35:46,659 --> 00:35:50,019 Pure kinematics. But in reality, robot arms are 755 00:35:50,019 --> 00:35:52,500 made of cast iron and aluminum. They weigh hundreds 756 00:35:52,500 --> 00:35:55,139 of pounds. To actually make them move at those 757 00:35:55,139 --> 00:35:58,139 velocities, we need real physical force. We must 758 00:35:58,139 --> 00:36:01,280 transition from kinematics to kinetics. We have 759 00:36:01,280 --> 00:36:04,469 to open chapter nine, dynamics and control. This 760 00:36:04,469 --> 00:36:06,670 is where the mathematical model collides with 761 00:36:06,670 --> 00:36:09,730 reality. If kinematics is the geometry of intent, 762 00:36:10,690 --> 00:36:13,250 dynamics is the physics of execution. How do 763 00:36:13,250 --> 00:36:16,250 we actually command a thousand pound arm to trace 764 00:36:16,250 --> 00:36:19,610 a perfect line? Quark starts with the historical 765 00:36:19,610 --> 00:36:22,969 baseline, independent joint control. It is the 766 00:36:22,969 --> 00:36:25,050 simplest and historically most common way to 767 00:36:25,050 --> 00:36:27,389 control a manipulator. You treat the robot not 768 00:36:27,389 --> 00:36:31,070 as a unified machine, but as six separate independent 769 00:36:31,070 --> 00:36:33,769 mechanisms. Like six independent motors that 770 00:36:33,769 --> 00:36:36,610 just happen to be bolted together. Exactly. You 771 00:36:36,610 --> 00:36:39,690 wrap a control loop, typically a PID loop, proportional 772 00:36:39,690 --> 00:36:42,429 integral derivative, around each individual motor. 773 00:36:43,250 --> 00:36:45,150 So you completely ignore the existence of the 774 00:36:45,150 --> 00:36:47,690 rest of the arm. Exactly. Joint 2 doesn't know 775 00:36:47,690 --> 00:36:50,469 Joint 3 exists. The controller just looks at 776 00:36:50,469 --> 00:36:52,610 the encoder error for Joint 2 and demands torque 777 00:36:52,610 --> 00:36:55,880 from the motor to shrink that error. But the 778 00:36:55,880 --> 00:36:57,960 physics of a manipulator arm makes that assumption 779 00:36:57,960 --> 00:37:00,079 fundamentally flawed, doesn't it? Because the 780 00:37:00,079 --> 00:37:02,579 load on joint two changes wildly depending on 781 00:37:02,579 --> 00:37:05,239 what joint three is doing. Dramatically. Let's 782 00:37:05,239 --> 00:37:08,119 look at gravity and inertia. If you have a two 783 00:37:08,119 --> 00:37:10,659 -link arm and it extends perfectly horizontally, 784 00:37:11,239 --> 00:37:13,179 the shoulder joint is suddenly holding up the 785 00:37:13,179 --> 00:37:15,480 entire mass of the arm at maximum mechanical 786 00:37:15,480 --> 00:37:18,699 leverage. The gravity load is immense. The PID 787 00:37:18,699 --> 00:37:20,920 controller on the shoulder has to output massive 788 00:37:20,920 --> 00:37:23,920 torque just to hold the arm steady. And if the 789 00:37:23,920 --> 00:37:26,619 elbow joint actuates and points the forearm straight 790 00:37:26,619 --> 00:37:28,780 down toward the floor, the center of mass shifts 791 00:37:28,780 --> 00:37:31,119 backward. The gravity load on the shoulder drops 792 00:37:31,119 --> 00:37:34,079 instantly. And it's even worse when the arm is 793 00:37:34,079 --> 00:37:36,579 moving. Inertia changes by the square of the 794 00:37:36,579 --> 00:37:38,900 radius. If the arm is tucked in tight to the 795 00:37:38,900 --> 00:37:41,639 base, the inertia tensor is small. The shoulder 796 00:37:41,639 --> 00:37:44,539 motor can accelerate the arm easily. But if the 797 00:37:44,539 --> 00:37:47,360 elbow extends the arm fully out, the inertia 798 00:37:47,360 --> 00:37:50,840 quadruples. Exactly. So if I spend a week perfectly 799 00:37:50,840 --> 00:37:53,159 tuning my PID gains for the shoulder motor while 800 00:37:53,159 --> 00:37:56,260 the arm is tucked in, making it snappy and responsive. 801 00:37:56,900 --> 00:37:59,340 The moment the arm extends, those exact same 802 00:37:59,340 --> 00:38:01,500 PID gains are now driving an inertia that is 803 00:38:01,500 --> 00:38:04,340 four times larger. The control loop becomes massively 804 00:38:04,340 --> 00:38:07,099 underdamped and the arm will start oscillating 805 00:38:07,099 --> 00:38:10,880 wildly. So the independent joint controller views 806 00:38:10,880 --> 00:38:13,340 all of these massive physical shifts, the change 807 00:38:13,340 --> 00:38:16,039 in gravity, the shifting inertia, as mysterious 808 00:38:16,039 --> 00:38:18,539 disturbances. It's a constantly playing catch 809 00:38:18,539 --> 00:38:21,480 -up against the physics of its own body. To achieve 810 00:38:21,480 --> 00:38:23,860 high -speed, high -precision motion, you cannot 811 00:38:23,860 --> 00:38:26,199 treat physics as a disturbance. You must model 812 00:38:26,199 --> 00:38:28,739 it. We have to write out the rigid body equations 813 00:38:28,739 --> 00:38:31,579 of motion. We are talking about mass matrices, 814 00:38:31,980 --> 00:38:36,059 centrifugal forces, Coriolis forces. Yes. Cork 815 00:38:36,059 --> 00:38:38,340 details the dynamic equations, which take the 816 00:38:38,340 --> 00:38:43,679 form of MQ double dot plus CQQ dot plus GQ equals 817 00:38:43,679 --> 00:38:47,929 tau. That's a mouthful. What does it do? It calculates 818 00:38:47,929 --> 00:38:50,690 exactly how much torque, tau, is required at 819 00:38:50,690 --> 00:38:53,030 every joint to achieve a specific joint acceleration, 820 00:38:53,269 --> 00:38:55,469 q double dot, given the current pose and the 821 00:38:55,469 --> 00:38:57,989 current velocity. Let's break down the most chaotic 822 00:38:57,989 --> 00:39:01,190 part of that equation, the C matrix, the dynamic 823 00:39:01,190 --> 00:39:04,550 coupling, centrifugal and Coriolis forces. When 824 00:39:04,550 --> 00:39:06,889 multiple links are moving simultaneously, they 825 00:39:06,889 --> 00:39:09,590 physically fight each other. Dynamic coupling 826 00:39:09,590 --> 00:39:12,420 is highly nonlinear. If the shoulder motor is 827 00:39:12,420 --> 00:39:14,739 accelerating rapidly, the physical force of that 828 00:39:14,739 --> 00:39:18,199 acceleration whips through the links and exerts 829 00:39:18,199 --> 00:39:21,010 a torque on the elbow joint. The elbow motor 830 00:39:21,010 --> 00:39:23,349 has to generate counter torque just to hold its 831 00:39:23,349 --> 00:39:26,550 position, let alone execute its own motion. Yes. 832 00:39:26,969 --> 00:39:29,329 So what does this all mean? I want to use an 833 00:39:29,329 --> 00:39:32,170 analogy to really cement how weird Coriolis forces 834 00:39:32,170 --> 00:39:35,570 are. Imagine riding on one of those old school 835 00:39:35,570 --> 00:39:38,630 spinning playground roundabouts. Oh, the ones 836 00:39:38,630 --> 00:39:41,349 responsible for so many broken arms in the 1990s. 837 00:39:41,670 --> 00:39:45,130 Exactly those. The roundabout is spinning incredibly 838 00:39:45,130 --> 00:39:48,869 fast. That spinning platform represents the robot's 839 00:39:48,869 --> 00:39:50,750 shoulder joint, sweeping the arm through the 840 00:39:50,750 --> 00:39:53,940 air. Now while it's spinning, you decide to walk 841 00:39:53,940 --> 00:39:56,300 from the outer edge of the roundabout straight 842 00:39:56,300 --> 00:39:59,699 in toward the center axis. Walking inward represents 843 00:39:59,699 --> 00:40:02,280 the robot's elbow bending, pulling the mass closer 844 00:40:02,280 --> 00:40:05,650 to the base. Right. Because you were on the outer 845 00:40:05,650 --> 00:40:08,269 edge, your body was traveling at a very high 846 00:40:08,269 --> 00:40:11,639 linear velocity. As you walk inward to a smaller 847 00:40:11,639 --> 00:40:14,039 radius, the floor beneath you is spinning at 848 00:40:14,039 --> 00:40:17,400 a slower linear velocity. You were carrying excess 849 00:40:17,400 --> 00:40:20,489 momentum. So as you try to walk straight inward, 850 00:40:20,610 --> 00:40:23,809 you feel this intense, invisible sideways force 851 00:40:23,809 --> 00:40:26,170 violently pushing you off your straight path. 852 00:40:26,550 --> 00:40:29,010 You have to lean hard against it just to keep 853 00:40:29,010 --> 00:40:31,789 walking straight. That is the Coriolis force. 854 00:40:32,110 --> 00:40:34,769 That's a perfect analogy. For a robot arm, if 855 00:40:34,769 --> 00:40:36,909 the shoulder is spinning the roundabout and the 856 00:40:36,909 --> 00:40:39,250 elbow starts to bend inward, walking to the center, 857 00:40:39,869 --> 00:40:42,969 a massive lateral Coriolis torque hits the elbow 858 00:40:42,969 --> 00:40:45,909 motor. And it depends on the speeds of both joints, 859 00:40:45,949 --> 00:40:48,789 right? Yes. It is dependent on the product of 860 00:40:48,789 --> 00:40:52,480 the velocities of multiple joints. If a simple 861 00:40:52,480 --> 00:40:54,900 PID controller encounters a strong Coriolis force, 862 00:40:55,320 --> 00:40:57,800 it will be physically overwhelmed, and the robot 863 00:40:57,800 --> 00:41:00,019 will deviate heavily from its intended path. 864 00:41:00,219 --> 00:41:03,139 So, how do we tame the roundabout? How do we 865 00:41:03,139 --> 00:41:05,179 get these machines to move with submillimeter 866 00:41:05,179 --> 00:41:08,239 precision at five meters per second? We use a 867 00:41:08,239 --> 00:41:11,360 paradigm called computed torque control. Quark 868 00:41:11,360 --> 00:41:13,820 spends a lot of time on this because it is the 869 00:41:13,820 --> 00:41:16,900 gold standard for industrial robotics. How does 870 00:41:16,900 --> 00:41:19,440 it work? Since we have mathematically modeled 871 00:41:19,440 --> 00:41:22,059 the equations of motion, we know the exact mass 872 00:41:22,059 --> 00:41:24,139 matrix, the exact gravity vectors, the exact 873 00:41:24,139 --> 00:41:26,980 Coriolis math. Based on our current speeds, we 874 00:41:26,980 --> 00:41:29,239 can calculate exactly what those disturbance 875 00:41:29,239 --> 00:41:32,579 torques will be at any given microsecond. Instead 876 00:41:32,579 --> 00:41:34,920 of reacting to the disturbance after it happens, 877 00:41:35,199 --> 00:41:37,659 we anticipate it. We don't just anticipate it, 878 00:41:37,860 --> 00:41:40,699 we completely cancel it. In computed torque control, 879 00:41:40,820 --> 00:41:43,519 the software computes the exact nonlinear physics 880 00:41:43,519 --> 00:41:46,300 forces, and it commands the motors to generate 881 00:41:46,300 --> 00:41:49,860 the exact opposite torques. Whoa. It mathematically 882 00:41:49,860 --> 00:41:53,699 injects a real -time anti -gravity, anti -inertia, 883 00:41:53,860 --> 00:41:56,380 anti -coriola signal into the motor drives. That 884 00:41:56,380 --> 00:41:59,300 is mind -blowing. You are using real -time calculus 885 00:41:59,300 --> 00:42:01,840 to perfectly erase the physical limitations of 886 00:42:01,840 --> 00:42:05,219 the machine. Mathematically, yes. By dynamically 887 00:42:05,219 --> 00:42:07,420 canceling out the messy non -linear physics, 888 00:42:07,960 --> 00:42:10,280 the robot mathematically appears in the underlying 889 00:42:10,280 --> 00:42:13,079 control system as a set of perfectly decoupled, 890 00:42:13,260 --> 00:42:15,900 idealized linear masses. And then you just use 891 00:42:15,900 --> 00:42:19,219 normal PID. You can grab a very aggressive, highly 892 00:42:19,219 --> 00:42:21,239 tuned linear controller around that idealized 893 00:42:21,239 --> 00:42:23,500 system to drive it with absolute perfection. 894 00:42:24,199 --> 00:42:27,139 But let me challenge that perfection. The math 895 00:42:27,139 --> 00:42:29,900 assumes you can just command a motor to produce 896 00:42:29,900 --> 00:42:32,699 50 Newton meters of torque and it instantly does 897 00:42:32,699 --> 00:42:35,460 it. But motors aren't equations. They are coils 898 00:42:35,460 --> 00:42:38,139 of copper wire and magnets. Torque outlines that 899 00:42:38,139 --> 00:42:40,880 true torque control is vastly harder than position 900 00:42:40,880 --> 00:42:43,139 control because of motor limits. You are hitting 901 00:42:43,139 --> 00:42:45,559 on the fundamental hardware barrier of computed 902 00:42:45,559 --> 00:42:49,030 torque control. The math says, give me 15 Newton 903 00:42:49,030 --> 00:42:52,659 meters right now. But an electric motor has physical 904 00:42:52,659 --> 00:42:56,500 limits. First, you have torque ripple. Right, 905 00:42:56,519 --> 00:42:59,739 because of the magnets. The magnetic field inside 906 00:42:59,739 --> 00:43:02,739 a motor isn't perfectly smooth. As the rotor 907 00:43:02,739 --> 00:43:05,380 passes the stator poles, the torque output ripples, 908 00:43:05,800 --> 00:43:08,559 introducing high frequency noise into your perfect 909 00:43:08,559 --> 00:43:11,179 mathematical cancellation. And you have current 910 00:43:11,179 --> 00:43:14,079 limits. You can't just push infinite amps into 911 00:43:14,079 --> 00:43:16,989 the copper coils. If you command massive torque 912 00:43:16,989 --> 00:43:19,829 to cancel a Coriolis force, the amplifier might 913 00:43:19,829 --> 00:43:22,710 hit its thermal limit and clip the current, meaning 914 00:43:22,710 --> 00:43:25,289 your computed torque is never actually delivered. 915 00:43:25,829 --> 00:43:27,809 The math asks for a check the hardware can't 916 00:43:27,809 --> 00:43:32,150 cache. And worse, you have back EMF. As the motor 917 00:43:32,150 --> 00:43:35,070 spins faster, it acts like a generator, creating 918 00:43:35,070 --> 00:43:37,550 a voltage that actively fights the voltage you 919 00:43:37,550 --> 00:43:40,320 are applying. At high speeds, the motor simply 920 00:43:40,320 --> 00:43:42,320 cannot draw enough current to generate the torque 921 00:43:42,320 --> 00:43:45,320 the math is demanding. So true torque control 922 00:43:45,320 --> 00:43:48,119 is a hardware nightmare. You need incredibly 923 00:43:48,119 --> 00:43:51,119 high quality, low friction gearboxes, high bandwidth 924 00:43:51,119 --> 00:43:53,320 current amplifiers, and perfect feed -forward 925 00:43:53,320 --> 00:43:56,599 friction models. Exactly. It is highly complex. 926 00:43:57,059 --> 00:43:59,360 But it is necessary for high performance, rigid 927 00:43:59,360 --> 00:44:02,699 control. But if we connect this to the bigger 928 00:44:02,699 --> 00:44:05,679 picture and look at the trajectory of robotics, 929 00:44:06,619 --> 00:44:08,320 computed torque control is still fundamentally 930 00:44:08,320 --> 00:44:12,219 focused on position. Right. It is about forcing 931 00:44:12,219 --> 00:44:15,019 the robot's joints to perfectly follow a predefined 932 00:44:15,019 --> 00:44:17,460 trajectory, aggressively rejecting any physical 933 00:44:17,460 --> 00:44:20,139 disturbances to stay on path. It's the heritage 934 00:44:20,139 --> 00:44:22,719 of the automotive assembly line. The robot is 935 00:44:22,719 --> 00:44:25,599 going to reach coordinate XYZ, and it doesn't 936 00:44:25,599 --> 00:44:28,360 care what is in its way. If a human steps in 937 00:44:28,360 --> 00:44:30,320 front of a position -controlled factory robot, 938 00:44:30,679 --> 00:44:32,579 the computed torque controller sees the human 939 00:44:32,579 --> 00:44:35,880 as an unmodeled disturbance mass. And it immediately 940 00:44:35,880 --> 00:44:38,119 commands maximum motor current to blast through 941 00:44:38,119 --> 00:44:40,539 the human and eliminate the position error. It 942 00:44:40,539 --> 00:44:43,760 is entirely unyielding. But modern robotics, 943 00:44:44,199 --> 00:44:47,000 collaborative robots, surgical robots, robots 944 00:44:47,000 --> 00:44:50,099 operating in domestic kitchens, requires interaction. 945 00:44:50,739 --> 00:44:52,840 We don't just want rigid position control. We 946 00:44:52,840 --> 00:44:56,000 want controlled compliance. We want to control 947 00:44:56,000 --> 00:44:58,840 the forces of interaction. And that brings Quark 948 00:44:58,840 --> 00:45:01,579 to task space dynamics, specifically operational 949 00:45:01,579 --> 00:45:04,699 space control. Operational space control. This 950 00:45:04,699 --> 00:45:06,940 is where we stop thinking about controlling joints 951 00:45:06,940 --> 00:45:10,239 and we directly control the physics at the tooltip. 952 00:45:10,289 --> 00:45:13,110 Yes. Usama Khatib formalized this, right? He 953 00:45:13,110 --> 00:45:15,449 did. Instead of writing the dynamic equations 954 00:45:15,449 --> 00:45:18,570 in terms of joint angles, we used the Jacobian 955 00:45:18,570 --> 00:45:21,110 matrix to project all of those mass matrices, 956 00:45:21,429 --> 00:45:24,409 Coriolis forces, and gravity vectors out to the 957 00:45:24,409 --> 00:45:27,690 end of vector. So mathematically, the robot stops 958 00:45:27,690 --> 00:45:30,849 being a complex arm of six links and abstracts 959 00:45:30,849 --> 00:45:33,269 itself into a single multi -dimensional mass 960 00:45:33,269 --> 00:45:36,750 located exactly at the tool tip. Precisely. And 961 00:45:36,750 --> 00:45:39,090 once the dynamics are expressed at the tooltip, 962 00:45:39,309 --> 00:45:42,250 we can apply impedance control. This is one of 963 00:45:42,250 --> 00:45:43,909 the most brilliant concepts in modern control 964 00:45:43,909 --> 00:45:46,769 theory. How so? Instead of telling the hand to 965 00:45:46,769 --> 00:45:49,349 aggressively hold a rigid coordinate, we program 966 00:45:49,349 --> 00:45:52,090 the controller so that the tooltip mathematically 967 00:45:52,090 --> 00:45:55,969 behaves like a spring mass damper system. A programmable 968 00:45:55,969 --> 00:45:58,969 virtual spring attached to the hand. Yes. And 969 00:45:58,969 --> 00:46:01,469 because it's software, we can set the stiffness 970 00:46:01,469 --> 00:46:03,590 of that spring differently for every spatial 971 00:46:03,590 --> 00:46:06,940 dimension. Give me an example. If we want a robot 972 00:46:06,940 --> 00:46:10,059 to wipe down a table with a sponge, we set the 973 00:46:10,059 --> 00:46:12,360 virtual spring to be very stiff in the X and 974 00:46:12,360 --> 00:46:15,480 Y directions, parallel to the table, so it precisely 975 00:46:15,480 --> 00:46:18,579 follows the wiping pattern. But in the Z direction, 976 00:46:18,800 --> 00:46:21,179 down toward the wood, we set the virtual spring 977 00:46:21,179 --> 00:46:23,619 to be very soft. So when the robot presses the 978 00:46:23,619 --> 00:46:25,940 sponge against the table, the force sensors on 979 00:46:25,940 --> 00:46:29,139 the wrist detect the reaction force. The software 980 00:46:29,139 --> 00:46:32,039 sees that force and the virtual soft spring in 981 00:46:32,039 --> 00:46:36,380 the Z axis compresses. the robot physically yields. 982 00:46:36,940 --> 00:46:39,179 It maintains a gentle, constant contact force 983 00:46:39,179 --> 00:46:41,619 instead of blindly driving the end effector through 984 00:46:41,619 --> 00:46:44,639 the tabletop. It allows a rigid machine to safely 985 00:46:44,639 --> 00:46:47,639 interact with a messy, geometrically uncertain 986 00:46:47,639 --> 00:46:50,539 environment. It feels the forces and responds 987 00:46:50,539 --> 00:46:53,719 compliantly. It transforms a dangerous industrial 988 00:46:53,719 --> 00:46:56,880 arm into a collaborative partner. But I have 989 00:46:56,880 --> 00:46:59,619 a serious safety concern here. Impedance control 990 00:46:59,619 --> 00:47:02,739 is a software trick. It relies on a force sensor 991 00:47:02,739 --> 00:47:05,230 reading an impact. Sending that data to a computer, 992 00:47:05,349 --> 00:47:07,510 the computer calculating the operational space 993 00:47:07,510 --> 00:47:09,909 dynamics and then commanding the motor drives 994 00:47:09,909 --> 00:47:12,929 to yield. That control loop takes time. It takes 995 00:47:12,929 --> 00:47:15,730 milliseconds. If the software crashes or if the 996 00:47:15,730 --> 00:47:18,329 arm impacts a human head at two meters per second, 997 00:47:18,869 --> 00:47:20,570 the kinetic energy is transferred before the 998 00:47:20,570 --> 00:47:22,989 software can even react. You have identified 999 00:47:22,989 --> 00:47:24,969 the ultimate limit of purely software -based 1000 00:47:24,969 --> 00:47:27,969 active compliance. High -frequency impacts cannot 1001 00:47:27,969 --> 00:47:30,320 be absorbed by software. They are governed by 1002 00:47:30,320 --> 00:47:32,420 the raw mechanical inertia of the links in the 1003 00:47:32,420 --> 00:47:35,380 gearbox. So what's the fix? That is exactly why 1004 00:47:35,380 --> 00:47:37,420 Quirk concludes this chapter by discussing the 1005 00:47:37,420 --> 00:47:40,739 hardware solution, series elastic actuators or 1006 00:47:40,739 --> 00:47:44,380 SEAs. Hardware compliance. We physically change 1007 00:47:44,380 --> 00:47:47,139 how the George is built. Yes. Pioneered by Gil 1008 00:47:47,139 --> 00:47:50,300 Pratt at MIT. Instead of rigidly connecting the 1009 00:47:50,300 --> 00:47:53,400 motor's gearbox to the robot's link, you insert 1010 00:47:53,400 --> 00:47:57,280 a physical calibrated torsion spring between 1011 00:47:57,280 --> 00:47:59,820 the gearbox output and the link. Wait, a real 1012 00:47:59,820 --> 00:48:02,619 spring? Yeah. So in the modal turns, it twists 1013 00:48:02,619 --> 00:48:05,460 a real steel spring, and then the spring pushes 1014 00:48:05,460 --> 00:48:08,750 the arm. Exactly. And you put a high resolution 1015 00:48:08,750 --> 00:48:10,949 encoder across that spring to measure exactly 1016 00:48:10,949 --> 00:48:15,210 how much it deflects. Using Hooke's law, F equals 1017 00:48:15,210 --> 00:48:19,489 kx, or tau equals k delta theta, that spring 1018 00:48:19,489 --> 00:48:22,130 deflection gives you an instantaneous, perfectly 1019 00:48:22,130 --> 00:48:24,929 clean measurement of the exact torque being applied 1020 00:48:24,929 --> 00:48:27,349 to the joint, completely bypassing the noise 1021 00:48:27,349 --> 00:48:30,219 and friction of the gearbox. But more importantly, 1022 00:48:30,659 --> 00:48:33,780 it provides mechanical safety. If a robot with 1023 00:48:33,780 --> 00:48:36,219 series elastic actuators accidentally hits a 1024 00:48:36,219 --> 00:48:39,800 human, that physical steel spring instantly compresses. 1025 00:48:40,039 --> 00:48:42,199 It mechanically absorbs the shock of the impact 1026 00:48:42,199 --> 00:48:44,519 before the software even knows an impact occurred. 1027 00:48:45,059 --> 00:48:47,760 It acts as a low -pass filter for physical shocks. 1028 00:48:48,500 --> 00:48:50,780 It guarantees physical safety in a way that software 1029 00:48:50,780 --> 00:48:54,309 alone never can. But... There is always a trade 1030 00:48:54,309 --> 00:48:56,469 -off in engineering. If you put a bouncy spring 1031 00:48:56,469 --> 00:48:58,750 in every joint of your robot, doesn't it become 1032 00:48:58,750 --> 00:49:02,150 a wobbly mess? How can you do precision manufacturing 1033 00:49:02,150 --> 00:49:05,750 with a robot made of springs? You lose high -frequency 1034 00:49:05,750 --> 00:49:08,730 position bandwidth. A series elastic robot will 1035 00:49:08,730 --> 00:49:11,530 never be as stiff or as precise at high speeds 1036 00:49:11,530 --> 00:49:15,570 as a rigid industrial arm. You trade pure positional 1037 00:49:15,570 --> 00:49:18,369 accuracy for high fidelity force control and 1038 00:49:18,369 --> 00:49:21,190 absolute safety. So it's a design choice based 1039 00:49:21,190 --> 00:49:24,230 on the application. It is. If you are milling 1040 00:49:24,230 --> 00:49:27,130 engine blocks, you want rigid joints and computed 1041 00:49:27,130 --> 00:49:29,969 torque control. If you are building a robot to 1042 00:49:29,969 --> 00:49:32,449 help nurses lift patients in a hospital, you 1043 00:49:32,449 --> 00:49:35,349 want series elastic actuators and impedance control. 1044 00:49:36,050 --> 00:49:38,329 Cork lays out the entire spectrum of these control 1045 00:49:38,329 --> 00:49:41,869 architectures perfectly. It is a phenomenal evolution. 1046 00:49:42,170 --> 00:49:44,769 I mean, we started this hour looking at the pure, 1047 00:49:45,030 --> 00:49:47,650 rigid geometry of a robotic arm. We mapped its 1048 00:49:47,650 --> 00:49:50,050 skeleton using Denevet -Hartenberg parameters 1049 00:49:50,050 --> 00:49:53,050 and pose graphs, exploring the algebraic elegance 1050 00:49:53,050 --> 00:49:55,369 and the industrial nightmares they cause. The 1051 00:49:55,369 --> 00:49:58,309 API mismatch of hardware. Exactly. We injected 1052 00:49:58,309 --> 00:50:01,710 time using the Jacobian matrix to translate joint 1053 00:50:01,710 --> 00:50:04,349 speeds into Cartesian velocities, navigating 1054 00:50:04,349 --> 00:50:06,710 the geometric black holes of singularities and 1055 00:50:06,710 --> 00:50:09,170 the magic of null space motion. And then we added 1056 00:50:09,170 --> 00:50:11,900 the physics. Right. And finally, we crashed into 1057 00:50:11,900 --> 00:50:14,619 reality. We added mass, gravity, and Coriolis 1058 00:50:14,619 --> 00:50:17,579 forces and saw how computed torque control mathematically 1059 00:50:17,579 --> 00:50:20,760 cancels out the physics, and ultimately how programmable 1060 00:50:20,760 --> 00:50:23,480 springs and series elastic actuators turn rigid 1061 00:50:23,480 --> 00:50:27,000 industrial machines into safe, compliant, interactive 1062 00:50:27,000 --> 00:50:29,920 systems. It truly highlights that a modern robotic 1063 00:50:29,920 --> 00:50:32,159 manipulator is a masterpiece of integration. 1064 00:50:33,090 --> 00:50:36,190 It is an incredible, inseparable fusion of pure, 1065 00:50:36,630 --> 00:50:39,949 abstract linear algebra and messy unforgiving 1066 00:50:39,949 --> 00:50:42,489 physical reality. It's a beautiful synthesis 1067 00:50:42,489 --> 00:50:45,510 of math and metal, but as we wrap up this deep 1068 00:50:45,510 --> 00:50:47,590 dive, I want to leave you, our listener, with 1069 00:50:47,590 --> 00:50:50,869 one final provocative thought. Let's hear it. 1070 00:50:51,150 --> 00:50:53,530 We have spent an hour talking about how a robot 1071 00:50:53,530 --> 00:50:56,329 can mathematically calculate a path, how it can 1072 00:50:56,329 --> 00:50:59,510 optimize its manipulability, and how it can gently 1073 00:50:59,510 --> 00:51:02,349 apply a precise force to a specific coordinate 1074 00:51:02,349 --> 00:51:05,739 in space. But we never actually asked how the 1075 00:51:05,739 --> 00:51:07,739 robot knows where that coordinate is in the first 1076 00:51:07,739 --> 00:51:10,440 place. A critical blind spot. All of our math 1077 00:51:10,440 --> 00:51:12,260 today assumed the coordinates of the target, 1078 00:51:12,659 --> 00:51:15,639 the coffee cup, the welding seam, were already 1079 00:51:15,639 --> 00:51:18,630 perfectly known and fed into the algorithm. But 1080 00:51:18,630 --> 00:51:21,730 in a real unstructured world, things aren't perfectly 1081 00:51:21,730 --> 00:51:25,090 pictured. Things move. Exactly. A robot with 1082 00:51:25,090 --> 00:51:27,710 perfect Jacobian mathematics and perfect series 1083 00:51:27,710 --> 00:51:30,070 elastic impedance control is still incredibly 1084 00:51:30,070 --> 00:51:32,789 dangerous if it's blind. It's just swinging a 1085 00:51:32,789 --> 00:51:35,050 perfectly compliant virtual spring around in 1086 00:51:35,050 --> 00:51:37,989 the dark. To truly interact with an unpredictable 1087 00:51:37,989 --> 00:51:40,110 world, our robot needs to be able to perceive 1088 00:51:40,110 --> 00:51:42,469 it. Which perfectly sets the stage for the next 1089 00:51:42,469 --> 00:51:44,769 phase of Cork's textbook. It absolutely does. 1090 00:51:45,190 --> 00:51:47,010 Join us next time for part four of this five 1091 00:51:47,010 --> 00:51:49,809 -part series, Computer Vision. We are finally 1092 00:51:49,809 --> 00:51:52,880 going to give our robot eyes. Until then... Keep 1093 00:51:52,880 --> 00:51:55,019 unpacking the math, keep questioning the physics, 1094 00:51:55,219 --> 00:51:57,139 and we will see you on the next deep dive.