🤗 lerobot-diffusion-pusht-sim

This model repository contains LeRobot Physical AI Simulation Policy & Rollout Artifacts for the PushT 2D block manipulation and ALOHA dual-arm simulation tasks.

Built using Hugging Face's LeRobot framework with Diffusion Policy (ResNet-18 Backbone) and Action Chunking Transformer (ACT).

---

📊 Benchmark Results

Environment	Policy Architecture	Observation Space	Action Space	Average Reward	Inference Speed
PushT 2D	Diffusion Policy (DDPM, 100 timesteps)	RGB (96, 96, 3) + 2D Pos (2,)	2D Target Coord (2,)	66.70 / 100	~1.90 FPS (MPS)
ALOHA Transfer	ACT (Action Chunking Transformer)	Top RGB (480, 640, 3) + 14-DoF State	14-DoF Joint Pos	8.00 (Success)	~45.0 FPS (MPS)

---

🎬 Rollout Demo Videos

• PushT 2D Manipulation: pusht_rollout_demo.mp4
• ALOHA Transfer Cube: aloha_transfer_cube_demo.mp4

---

🚀 How to Run Locally

1. Install Dependencies

pip install lerobot "gym-pusht>=0.1.5" "gym-aloha>=0.1.4" diffusers torch

2. Run Simulation Rollout

# PushT Diffusion Policy
python run_simulation.py --env pusht --episodes 1 --max-steps 150

# ALOHA Dual-Arm ACT Policy
python run_simulation.py --env aloha_transfer --episodes 1 --max-steps 150

3. Interactive Mode (Keyboard Control)

python interactive_sim.py --env pusht