Triton Droids Humanoid Robot Simulator

Official repository for Triton Droids simulations.

This codebase provides a framework to train, evaluate, and test reinforcement learning (RL) policies using MJX and MuJoCo.

Installation and Configuration

Please refer to [setup.md] for installation and configuration steps.

🚀 Features

This simulator codebase provides a platform for:

• Train: Learn policies using MJX and Brax PPO.

• Play: Test trained policies inside MuJoCo environments.

• Sim2Sim Transfer: Transfer policies between simulators to improve robustness. (WIP)

• Sim2Real Transfer: Deploy policies to physical robots for real-world control. (Future goal)

📂 Project Structure

/simulation
├── /scripts                              # Shell scripts to automate training execution
├── /src
│   ├── /configs                          # Dataclass config files for environmnets, sim, and rl params
│   ├── /locomotion                       # (mjx) Locomotion task environments
│   ├── /rewards                          # Reward functions
│   ├── /robots                           # Robot object to represent MuJoCo model
│   ├── /scripts                          # Python entry points for training and policy playback
│   ├── /sim                              # MuJoCo utilities (physics rendering and sim-state management)
│   ├── /tools                            # Experiment tools
│   ├── /utils                            # General-purpose uitilites

Usage

1. Training

Run the following command to start training:

./scripts/single_train_mjx.sh

Parameter Description

• VIDEO: Flag to enable/disable video recording during training.

• VIDEO_LENGTH: Number of simulation steps (frames) to record per video. Controls how long each video lasts.

• VIDEO_INTERVAL:Interval (in training steps or timesteps) between video recordings.

2. Play

Coming soon.

3. Sim2Sim

Coming soon.

Future Plans

• Ros Integration
• Isaac Lab

Acknowledgments

This repository incorporates code from toddlerbot.