Day 26
Isaac Lab: a peek at the production-grade alternative
This is a valid v1.0 placeholder page for the later curriculum arc. Full interactive lab treatment ships after Week 1 dogfooding.
LECTURE & READING
Glossary primer (10 min)
- Isaac Lab — NVIDIA's GPU-accelerated Core ConceptsRobotA physical system with sensors and actuators that can observe the world and take actions. Simulation & Sim-to-RealSimulationA virtual environment where robots can be trained or tested. framework, built on Isaac Sim (Omniverse). Successor to Isaac Gym (deprecated 2024). Industry standard for humanoid/dexterous research in 2026.
- Isaac Sim — NVIDIA's full-fidelity simulator. Photorealistic rendering, USD scene format, physics powered by PhysX.
- USD (Universal Scene Description) — Pixar's scene-graph format. Standard for Isaac Sim assets.
- GR00T integration — Isaac Lab ships with NVIDIA's GR00T humanoid env reference implementations.
- rsl_Imitation & Reinforcement LearningReinforcement Learning (RL)Teaching a robot through trial and error using rewards. — ETH Zurich's Imitation & Reinforcement LearningReinforcement Learning (RL)Teaching a robot through trial and error using rewards. library. The de-facto trainer for Isaac Lab.
- Skrl — Alternative Imitation & Reinforcement LearningReinforcement Learning (RL)Teaching a robot through trial and error using rewards. library used in Isaac Lab. Multi-agent capable.
Real-world analogy
MuJoCo Playground (Days 23–25) is "lightweight, fast, indie". Isaac Lab is "AAA studio, higher fidelity, bigger toolchain, longer install time". They cover the same use cases; choosing depends on your project and team.
Hour 1 — Setup Isaac Lab
This is annoying but unavoidable.
# Provision an Nebius instance with at least 100 GB disk, RTX/H100 GPU.
# Isaac Sim only runs on Linux with NVIDIA GPUs.
ssh -i ~/.ssh/nebius_key ubuntu@<your-ip>
# Install Isaac Sim 4.5+ via Omniverse Launcher (or Docker)
docker pull nvcr.io/nvidia/isaac-lab:2.2.0 # current as of Apr 2026
docker run --gpus all --rm -it -v $(pwd):/workspace nvcr.io/nvidia/isaac-lab:2.2.0 bash(In container)
cd /opt/IsaacLab
./isaaclab.sh --versionExpected: Isaac Lab 2.2.0.
Hour 2 — Run the official quadruped tutorial
./isaaclab.sh -p source/standalone/tutorials/03_envs/run_random_agent.py \
--task Isaac-Velocity-Flat-Anymal-C-v0 --headless --num_envs 64 --steps 200Expected: ANYmal-C robots (64 of them) appear in headless sim. Random actions; logs print env stats.
Then train PPO on the same env:
./isaaclab.sh -p source/standalone/workflows/rsl_rl/train.py \
--task Isaac-Velocity-Flat-Anymal-C-v0 --headless --num_envs 4096 --max_iterations 1500Expected: Robot LearningTrainingThe process of fitting a model using data or experience. takes ~20 minutes. Final mean Imitation & Reinforcement LearningRewardA score that tells the robot how well it is doing. ≈ 30. (ANYmal-C is similar mass class to Go1.)
LAB
Hour 3 — Lab: replicate Day 24's Go1 walk in Isaac Lab (75 min)
What you're building. Train ANYmal-C (Isaac Lab's default quadruped) in flat terrain via Isaac Lab's rsl_rl workflow. Side-by-side with your Day 24 Go1 result. Compare wall-clock, Evaluation & ResearchThroughputHow much data or how many actions a system can process in a given time., and final Imitation & Reinforcement LearningRewardA score that tells the robot how well it is doing..
Step 1 — Run the train command above (60 min)
Already started above. While it runs, watch the TensorBoard:
tensorboard --logdir=logs/rsl_rl --port=6006Full source continues in the committed curriculum files. The v1.0 page exposes the day flow and lab surface without inventing content.