MetaWorld Agent Migration (PPO → SAC + DrQ-v2)

Installation

Prerequisites

• Python 3.12+
• MuJoCo (for MetaWorld environments)

Quick Start

# Install with uv (recommended)
uv venv --python python 3.12
source .venv/bin/activate
uv sync

Training a Model

⚠️ The legacy PPO trainer has been removed. A new SAC + DrQ-v2 training pipeline is under construction and will land in an upcoming update.

Local Evaluation

# Evaluate on reach task
python evaluation.py --task <task_name> --episodes <num_episodes> --render-mode <human|rgb_array> --model-path <path_to_trained_model>

Available tasks:

reach-v3
push-v3
pick-place-v3
door-open-v3
door-close-v3
drawer-open-v3
button-press-topdown-v3
button-press-v3
peg-insert-side-v3
lever-pull-v3

Roadmap

• ✅ Enable multi-view pixel observations (corner, corner2, topview) alongside proprioception and task one-hot inputs.
• 🚧 Replace the PPO baseline with a SAC + DrQ-v2 implementation tuned for MetaWorld MT10 multi-task learning.
• 🔜 Reintroduce end-to-end training and evaluation documentation once the new agent lands.

TensorBoard Integration

# Training automatically starts TensorBoard
python train.py

# View logs at http://localhost:6006

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• MetaWorld - Robotics benchmark suite
• Metaworld Algorithms - Base algorithm implementations
• Kinitro - Agent submission framework
• JAX - High-performance machine learning
• Flax - Neural network library