Skip to content

Robot-Nav/RPO_IsaacLab

BranchesTags

Repository files navigation

RPO_IsaacLab

IsaacSim Isaac Lab RSL_RL Python Linux License

English | 中文 | Technical Docs

RPO_IsaacLab is a reinforcement learning training workspace for the RPO bipedal robot locomotion, built on NVIDIA Isaac Lab and RSL-RL. It provides a complete pipeline from motion retargeting to sim-to-sim validation and real-robot deployment — supporting walking, dancing, parkour, fall recovery, and more human-style locomotion skills.

Demo video： dance_play：

dance_play.mov

dance_sim2sim:

dance_sim2sim.mov

1. Features

  • AMP (Adversarial Motion Priors) — Stylized locomotion learning with discriminator-guided style rewards (walking, dancing)
  • BeyondMimic — Reference trajectory imitation with fall recovery
  • Parkour — Complex terrain traversal
  • Attention Encoder — Adaptive locomotion with terrain-aware attention
  • Interrupt Recovery — Robust recovery from external perturbations
  • GMR Motion Retargeting — BVH/FBX to robot-specific motion data pipeline
  • MuJoCo Sim2Sim — Policy transfer validation in MuJoCo
  • Atom01 Deployment — ONNX export and ROS2 inference for real robot deployment

2. Supported Tasks

Task Description
RPO-AMP / RPO-AMP-Play AMP walking style training / inference
RPO-AMP-Dance / RPO-AMP-Dance-Play AMP dance style training (LAFAN1 dataset)
RPO-BeyondMimic Reference trajectory imitation + fall recovery
RPO-Getup-Mimic Get-up motion learning
RPO-Parkour / RPO-Parkour-Play Parkour over complex terrain
RPO-Flat Flat terrain locomotion
RPO-Rough Rough terrain locomotion
RPO-AttnEnc Attention encoder training
RPO-Interrupt Interrupt recovery training

3. Installation

3.1 Prerequisites

  • Python 3.11
  • Isaac Sim 5.1.0
  • Isaac Lab v2.3
  • RSL-RL 3.3.0
  • CUDA 12.8+ (Blackwell GPU requires cu128 torch build)
  • Ubuntu 22.04 x64
  • NVIDIA Driver >= 535

3.2 Quick Start

# Clone the repository
git clone --recursive https://github.com/<your-username>/RPO_IsaacLab.git
cd RPO_IsaacLab

# Create conda environment
conda create -n rpo_isaaclab python=3.11 -y
conda activate rpo_isaaclab

# Install Isaac Sim 5.1
pip install torch==2.7.0 torchvision==0.22.0 --index-url https://download.pytorch.org/whl/cu128
pip install "isaacsim[all,extscache]==5.1.0" --extra-index-url https://pypi.nvidia.com

# Install Isaac Lab extensions (from IsaacLab v2.3 source)
cd /path/to/IsaacLab_RPO
./isaaclab.sh --install none

# Install project dependencies
cd RPO_IsaacLab
pip install -e ./rsl_rl
pip install -e ./robolab

Verify installation:

python robolab/scripts/tools/list_envs.py

4. Usage

4.1 Environment Verification

python robolab/scripts/tools/list_envs.py

4.2 Training

All tasks share the same entry point robolab/scripts/rsl_rl/train.py. Logs are saved to logs/rsl_rl/<experiment_name>/<timestamp>/, with a separate directory per task (e.g., rpo_amp, rpo_amp_dance).

# AMP walking (flat terrain)
python robolab/scripts/rsl_rl/train.py --task=RPO-AMP --headless --num_envs=4096

# AMP walking (rough terrain)
python robolab/scripts/rsl_rl/train.py --task=RPO-AMP-Rough --headless --num_envs=4096

# AMP dancing (LAFAN1)
python robolab/scripts/rsl_rl/train.py --task=RPO-AMP-Dance --headless --num_envs=2048

# AMP dancing (single motion)
python robolab/scripts/rsl_rl/train.py --task=RPO-AMP-Dance-Single --headless --num_envs=2048

# BeyondMimic reference tracking + fall recovery
python robolab/scripts/rsl_rl/train.py --task=RPO-BeyondMimic --headless --num_envs=4096

# Getup-Mimic
python robolab/scripts/rsl_rl/train.py --task=RPO-Getup-Mimic --headless --num_envs=4096

# Parkour
python robolab/scripts/rsl_rl/train.py --task=RPO-Parkour --headless --num_envs=4096

# Direct RL flat terrain
python robolab/scripts/rsl_rl/train.py --task=RPO-Flat --headless --num_envs=4096

# Direct RL rough terrain
python robolab/scripts/rsl_rl/train.py --task=RPO-Rough --headless --num_envs=4096

# Attention encoder
python robolab/scripts/rsl_rl/train.py --task=RPO-AttnEnc --headless --num_envs=4096

# Interrupt recovery
python robolab/scripts/rsl_rl/train.py --task=RPO-Interrupt --headless --num_envs=4096

Common options: --max_iterations <N> overrides the default iteration limit; --resume --load_run=<dir> resumes from a checkpoint; --logger=tensorboard enables TensorBoard; --distributed enables multi-GPU training (or launch via torchrun).

4.3 Testing / Playback

# AMP walking
python robolab/scripts/rsl_rl/play_amp.py --task=RPO-AMP-Play --num_envs=1

# AMP rough terrain
python robolab/scripts/rsl_rl/play_amp.py --task=RPO-AMP-Rough-Play --num_envs=1

# AMP dancing
python robolab/scripts/rsl_rl/play_amp.py --task=RPO-AMP-Dance-Play --num_envs=1

# AMP dancing (single motion)
python robolab/scripts/rsl_rl/play_amp.py --task=RPO-AMP-Dance-Single-Play --num_envs=1

# BeyondMimic
python robolab/scripts/rsl_rl/play_bm.py --task=RPO-BeyondMimic --num_envs=1

# Getup-Mimic
python robolab/scripts/rsl_rl/play_bm.py --task=RPO-Getup-Mimic --num_envs=1

# Direct RL (Flat / Rough / AttnEnc / Interrupt)
python robolab/scripts/rsl_rl/play.py --task=RPO-Flat --num_envs=1
python robolab/scripts/rsl_rl/play.py --task=RPO-Rough --num_envs=1
python robolab/scripts/rsl_rl/play.py --task=RPO-AttnEnc --num_envs=1
python robolab/scripts/rsl_rl/play.py --task=RPO-Interrupt --num_envs=1

# Parkour (add --exportonnx to export ONNX)
python robolab/scripts/rsl_rl/play_parkour.py --task=RPO-Parkour-Play --num_envs=1 --exportonnx

Use --load_run=<dir> to specify the training log directory, or --checkpoint=<path> to load a specific checkpoint. Playback automatically exports JIT/ONNX models to the exported/ directory.

4.4 Sim2Sim Validation

# Direct RL (Flat / Rough / AttnEnc / Interrupt)
python robolab/scripts/mujoco/sim2sim_rpo.py --load_model <exported/policy.pt>
python robolab/scripts/mujoco/sim2sim_rpo.py --load_model <path> --terrain
python robolab/scripts/mujoco/sim2sim_rpo_attn_enc.py --load_model <path>
python robolab/scripts/mujoco/sim2sim_rpo_interrupt.py --load_model <path>

# AMP / AMP-Rough (requires --terrain to load the full MJCF)
python robolab/scripts/mujoco/sim2sim_rpo_amp.py --load_model <path> --terrain

# BeyondMimic (load reference motion NPZ)
python robolab/scripts/mujoco/sim2sim_rpo_bm.py \
    --load_model <path> --motion_file <motion.npz>

# Parkour (separate ONNX graphs: encoder + actor)
python robolab/scripts/mujoco/sim2sim_rpo_parkour.py \
    --depth_encoder <0-depth_encoder.onnx> --actor <actor.onnx>

# Motion CSV visualization
python robolab/scripts/mujoco/play_motion_csv.py --motion_file <motion.csv>

Add --headless to disable the GUI and record a video. --load_model should point to the TorchScript exported/policy.pt, not the training checkpoint model_*.pt.

4.5 Motion Data Preparation

AMP and BeyondMimic require motion data in .pkl format. Use the GMR tool to retarget BVH/FBX mocap data to the RPO robot skeleton, then reorder joints for Isaac Lab:

# Batch GMR -> Isaac Lab retargeting
python robolab/scripts/tools/retarget/dataset_retarget.py \
    --robot rpo \
    --input_dir robolab/data/motions/rpo_gmr \
    --output_dir robolab/data/motions/rpo_lab \
    --config_file robolab/scripts/tools/retarget/config/rpo.yaml

# Single-file retargeting (supports --frame_range)
python robolab/scripts/tools/retarget/single_retarget.py \
    --robot rpo \
    --input_file <input.pkl> \
    --output_file <output.pkl> \
    --config_file robolab/scripts/tools/retarget/config/rpo.yaml

# BeyondMimic NPZ motion replay (visualization)
python robolab/scripts/tools/beyondmimic/replay_npz.py -f <motion.npz>

# CSV -> NPZ conversion
python robolab/scripts/tools/beyondmimic/csv_to_npz.py -f <input.csv> --input_fps 60

4.6 Robot Deployment

See atom01_deploy/README_CN.md for ONNX export, ROS2 inference node setup, and real robot deployment instructions.

5. Repository Structure

RPO_IsaacLab/
├── robolab/                    # Isaac Lab extensions (environments, assets, scripts)
│   ├── robolab/
│   │   ├── assets/robots/      # RPO robot definition
│   │   ├── tasks/              # RL environments (AMP, BeyondMimic, Parkour, etc.)
│   │   ├── utils/              # Math, buffers, noise, warp
│   │   └── scripts/            # Train, play, retarget, mujoco
│   └── data/motions/           # Motion datasets (.pkl)
├── rsl_rl/                     # RSL-RL 3.3.0 (PPO, AMP, symmetry)
├── atom01_deploy/              # Real robot deployment (ROS2, ONNX)
├── external/                   # GMR retargeting tool, dance data
└── TECHNICAL_DOC_CN.md         # Full technical documentation (Chinese)

6. FAQ

  • Empty robolab/rsl_rl directories: Run git submodule update --init --recursive
  • Isaac Lab imports not found: Activate the correct Python environment first
  • Task name not found: Run python robolab/scripts/tools/list_envs.py for the actual task IDs
  • Blackwell GPU (RTX 50xx) CUDA errors: Ensure torch is built for CUDA 12.8: python -c "import torch;print(torch.version.cuda)"

7. References

  • IsaacLab — NVIDIA robot simulation framework
  • rsl_rl — Legged robot RL library
  • RoboParty — Open-source robot learning projects
  • legged_gym — Legged robot training environments
  • legged_lab — Isaac Lab legged training extension
  • robot_lab — Isaac Lab robot training framework
  • InstinctLab — Instinctive policy framework

AMP algorithm from Adversarial Motion Priors for Stylized Locomotion (Peng et al., SIGGRAPH 2021).

Maintainer: Robot-Nav   |   Support: GitHub Issues

About

RPO_IsaacLab是面向 RPO 双足机器人的强化学习运动控制训练平台，基于 NVIDIA Isaac Lab 和 RSL-RL 构建。提供从动作捕捉重定向到策略仿真验证、再到实机部署的完整管线——支持行走、舞蹈、跑酷、跌倒恢复等多种类人运动技能训练。

blog.csdn.net/qq_56908984/article/details/162936619?spm=1011.2124.3001.6209

Topics

ampbeyondmimicppo

Resources

Readme
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors

Languages