This repository contains the standalone, lean Manager-Based IsaacLab implementation for the Aldebaran NAO humanoid robot locomotion environments.
IsaacLab (and its PhysX / rsl-rl backend) runs on a CUDA GPU. This repository is tested and verified with:
| Component | Version |
|---|---|
| OS | Ubuntu 24.04 LTS |
| Python | 3.11.15 |
| Isaac Sim | 5.1.0 |
| IsaacLab | 2.3.2 |
| rsl-rl-lib | 5.0.1 |
| PyTorch | 2.7.0 (CUDA 12.8) |
| Gymnasium | 1.2.1 |
Install IsaacLab by following its official documentation: https://isaac-sim.github.io/IsaacLab
Register the package into your Python / Isaac Sim environment by running:
cd nao_rl_lab
pip install -e .The following environments are registered via gymnasium and available for training:
Nao-Velocity-Flat-v0: Tuned omnidirectional flat-ground locomotion environment with push perturbations, friction randomization, and yaw-frame velocity tracking.Nao-Velocity-Flat-Play-v0: Evaluation variant of flat ground optimized for visualization and deployment benchmarking.Nao-Velocity-Rough-v0: Locomotion on gentle rough terrain (mild roughness, gentle slopes, and flat patches).Nao-Velocity-Rough-Play-v0: Evaluation variant for rough terrain visualization.
Train the locomotion policy using RSL-RL:
# Flat ground training
python scripts/train.py --task Nao-Velocity-Flat-v0 --num_envs 4096 --max_iterations 3000
# Rough terrain training
python scripts/train.py --task Nao-Velocity-Rough-v0 --num_envs 4096 --max_iterations 3000Visualize a trained policy directly inside Isaac Sim:
# Playback flat ground policy
python scripts/play.py --task Nao-Velocity-Flat-Play-v0 --checkpoint logs/rsl_rl/nao_tuned_flat/<RUN_TIMESTAMP>/model_2100.pt
# Playback rough terrain policy
python scripts/play.py --task Nao-Velocity-Rough-Play-v0 --checkpoint logs/rsl_rl/nao_rough/<RUN_TIMESTAMP>/model_3000.pt