babymore87
MyBoerse.bz Pro Member
Quadruped Robot Locomotion Train a Go2 to Walk with RL
Published 9/2026
MP4 | Video: h264, 1920x1080 | Audio: AAC, 44.1 KHz, 2 Ch
Language: English | Duration: 6h 10m | Size: 6.68 GB
Build gaits, train an RL policy in Isaac Lab, and learn to tell real walking from a policy that only looks like it
What you'll learn
Measure whether a robot is really walking: foot slip, duty factor, stride, clearance and diagonality, not vibes
Build hand-written gaits from one phase table, and find the point where they stop being the right tool
Train a quadruped locomotion policy in Isaac Lab, from task and observations to reward and terminations
Recognise reward hacking that scores 1.432 of 1.5 while the robot never lifts a foot
Write reward terms that produce a structurally different gait, and prove which ones were load-bearing by ablation
Transfer a policy from Isaac Lab to MuJoCo and prove the transfer, including the joint-order convention that silently breaks it
Test robustness properly: rough terrain, pushes, payload, centre-of-mass shift, latency and curriculum collapse
Audit a dynamic-motion claim against ballistics and actuator limits, and retract it when the numbers disagree
Export a trained policy to TorchScript and ONNX and write a deployment loop with no simulator in it
Read a metric that looks healthy and find the reason it is lying to you
Requirements
Comfortable Python: functions, classes, numpy arrays, reading a traceback
Basic linear algebra and calculus: vectors, matrices, derivatives. The kinematics is derived from scratch
A machine that can run MuJoCo. Most of the course is measurable on CPU
An NVIDIA GPU is needed to repeat the Isaac Lab training runs, but every training result ships as an artifact you can inspect without one
No prior reinforcement learning experience needed
Description
This course contains the use of artificial intelligence.
Legs are the easiest thing in robotics to animate, which is why a quadruped is wherelooks like walking andis walking come apart hardest. This course is one continuous argument built on that gap, and every step of it is measured rather than asserted.
It starts with a controller that fools you. A sine wave reads as a convincing trot from every camera angle. Then we fix every real bug in it, and the robot travels further while foot slip gets steadily worse. That is the point at which hand-written gaits are the wrong tool.
So we train a policy in Isaac Lab, and it cheats. It reaches 1.432 of a 1.5 maximum velocity-tracking reward and never lifts a foot: duty factor 97 to 98 percent, body height 0.116 m against a 0.27 m stand. It crawls to the commanded speed. The reward curve looks perfect. You cannot detect this failure using the reward.
The fix is the reward, not the algorithm. Two new terms later the same setup gives foot clearance 0 to 112 mm, duty 54 percent, height 0.375 m, and 1819 mm of travel. Then a six-run ablation removes one reward term at a time and shows which ones were structural. Four of my six predictions were wrong, and the lecture ships the scorecard.
Then robustness, and honest limits. Rough terrain, pushes, payload, centre-of-mass shifts, latency and a curriculum that scores worse while being right. The dynamic-motion section is built around a retraction: an internal note claimed a 378.5 degree backflip, and the audit shows it fails ballistics, actuators and orientation at once. The honest number is 80.4 degrees, and the lecture explains exactly which broken metric produced the first one.
You finish with something deployable. The policy exported to TorchScript and ONNX, and a runtime loop whose only imports are numpy and onnxruntime, with a self-test that asserts no simulator is present.
80 lectures. Every failure shown is a bug that actually happened during the build, and every number ships with the script that computed it.
Who this course is for
Robotics engineers who can train a policy but cannot yet tell a good gait from a policy that games the reward
ML engineers moving into physical systems who want a problem where the metric and the truth come apart
Simulation and controls engineers who want the MuJoCo and Isaac Lab paths compared on the same robot
Anyone who has watched a legged robot demo and wondered how much of it was real
Not for you if you want a plug-and-play walking policy with no measurement work
Homepage
Recommend Download Link Hight Speed | Please Say Thanks Keep Topic Live
No Password - Links are Interchangeable
