NVIDIA Isaac Lab 3.0 Custom Robots & RL Environments

Import a robot Isaac Lab has never seen, build its RL environment from scratch, train it with PPO, and export the policy
What you'll learn

---

Configure a robot Isaac Lab does not ship: URDF to USD, collision geometry, actuators, joint limits and drive gains, written from scratch
Convert a URDF without silently losing all 25 collision shapes, which the stock importer does while exiting zero
Build a complete manager-based RL environment: observations, rewards, terminations, events and curriculum
Assemble a 48-number observation vector one term at a time, and know why projected gravity is used instead of a quaternion
Train with PPO and rsl-rl at thousands of parallel environments on a single 6 GB laptop GPU
Read a learning curve honestly, and recognise a reward specification failure from the curve alone
Measure parallel scaling on YOUR hardware instead of trusting a number from a slide
Tell verification from validation, and say what a simulation result does not establish
Evaluate a trained policy with survival-conditioned metrics that a fallen robot cannot game
Export to ONNX and write a runtime that imports nothing from Isaac Lab
Requirements

---

A CUDA GPU. The whole course is built and measured on an RTX 4050 6 GB laptop card, so 6 GB is demonstrated as enough.
Comfortable Python: you should be able to read a dataclass and a config object without help.
No prior Isaac Lab experience. The install, the launcher and the directory layout are covered.
No prior reinforcement learning experience. The maths you need is derived in Section 2 and coded in Section 3.
Description

---

This course contains the use of artificial intelligence.
Most Isaac Lab material runs an example NVIDIA already shipped. You launch a task that exists, watch a robot that already has a config file, and learn nothing about what happens when the robot is yours.
This course takes the opposite path. The robot is a Google Barkour vB quadruped, 12 degrees of freedom, and Isaac Lab does not ship it: there is no robot config file for it, no registered task, nothing to import. Every configuration is written from an empty file.
It is measured, not asserted. The parallel-simulation benchmark is a real 24-run sweep on the build machine, repeated three times because a laptop under thermal management is not a clean instrument. One environment gives about 73 environment-steps per second; 4096 gives about 101,000, which is 1386x. You will also see where that curve stops being trustworthy, and why running the sweep once would have taught you noise.
It shows what breaks. The stock URDF importer drops every one of the robot's 25 collision elements, exits zero, and warns about none of it - the robot falls 19.87 m through the floor, and you watch it happen before you learn the fix. An effort_limit set on an implicit actuator is accepted and silently ignored. Drive gains arrive shrunk by the radians-to-degrees factor. A knee whose range is positive-only clamps a negative target to zero. Each of these is demonstrated, measured, and then fixed.
It is honest about scope. This is a simulation course. It trains a locomotion policy and exports it to ONNX; it does not deploy to physical hardware, and it says so rather than implying otherwise.
By the last section you will have taken a robot from a URDF file nobody had configured, through a working RL environment, to a trained policy that walks - and you will know which numbers you can trust on your own machine.
Who this course is for

---

Robotics engineers who need to bring a specific robot into Isaac Lab and have found the documentation assumes the robot is already supported
RL practitioners moving from Gym-style environments to GPU-parallel simulation
Graduate students and researchers who need a locomotion baseline they understand end to end rather than a repo they cloned
Anyone who has run an Isaac Lab example successfully and then had no idea how to change it
Not for you if you want a click-through tour of the GUI, or a course that deploys to real hardware
Homepage

---