Guguji Isaac Lab

Isaac Lab Extension

Modern GPU-parallel locomotion training for the Guguji biped

A focused Isaac Lab extension for training, evaluating, and exporting locomotion policies on the Guguji robot. The project migrates the workflow from Gazebo + ROS 2 + Stable-Baselines3 to Isaac Lab + RSL-RL for faster iteration and cleaner sim-to-real training loops.

Get started Training guide GitHub
Isaac Sim 4.5.0 Isaac Lab 2.1.0 Python 3.10 MIT License
Guguji flat terrain locomotion

Why this project

Parallel training

Run thousands of environments in parallel on GPU instead of iterating one-by-one in a traditional Gazebo setup.

Residual gait policy

The policy predicts residual joint positions on top of a sinusoidal reference gait, reducing exploration burden and improving convergence speed.

Built for deployment

Evaluation exports TorchScript and ONNX artifacts so the trained policy can move cleanly into downstream deployment workflows.

At a glance

Gym ID Terrain Envs Use
Isaac-Velocity-Flat-Guguji-v0 Flat 4096 Main training entry point
Isaac-Velocity-Flat-Guguji-Play-v0 Flat 50 Evaluation and visualization
Isaac-Velocity-Rough-Guguji-v0 Rough 2048 Terrain generalization training
Isaac-Velocity-Rough-Guguji-Play-v0 Rough 50 Rough-terrain evaluation

Migration context

Legacy setup This repository
Simulator Gazebo Fortress Isaac Sim / PhysX 5
RL stack Stable-Baselines3 RSL-RL
Parallel environments 1 4096
Curriculum Manual multi-stage script Native curriculum manager
Export .zip .pt TorchScript / .onnx

Quick start

cd ~/rlgpu_ws/IsaacLab
./isaaclab.sh -p -m pip install -e ~/Desktop/guguji_simulation/guguji_isaaclab/source/guguji_locomotion
./isaaclab.sh -p ~/Desktop/guguji_simulation/guguji_isaaclab/scripts/list_envs.py

cd ~/rlgpu_ws/IsaacLab
./isaaclab.sh -p ~/Desktop/guguji_simulation/guguji_isaaclab/scripts/rsl_rl/train.py \
  --task=Isaac-Velocity-Flat-Guguji-v0 \
  --num_envs=4096 \
  --headless

What you will find in the docs

  • Getting Started: installation, environment setup, and repository layout.
  • Training & Evaluation: standard train / play / export workflows.
  • Design Notes: reward shaping, reference gait design, and curriculum highlights.
  • Changelog: recent fixes and gait-quality improvements.

Repository structure

guguji_isaaclab/
├── scripts/
│   ├── list_envs.py
│   └── rsl_rl/
│       ├── train.py
│       ├── play.py
│       └── cli_args.py
└── source/guguji_locomotion/
    └── guguji_locomotion/
        ├── assets/
        └── tasks/locomotion/velocity/
            ├── velocity_env_cfg.py
            ├── mdp/
            └── config/guguji/