masondx/diffusion-aloha-insertion-joint-20k-bs32

Joint Diffusion Policy — ALOHA Insertion

EMA inference checkpoint for a standard joint Diffusion Policy trained on lerobot/aloha_sim_insertion_human. One policy receives the complete 14-dimensional bimanual state and jointly predicts the complete 14-dimensional action.

Configuration

• Training steps: 20,000
• Batch size: 32
• Training seed: 1000
• Policy type: diffusion
• State: 14-dimensional joint state
• Action: 14-dimensional joint action
• Visual input: observation.images.top at 480 x 640
• Observation steps: 2
• Prediction horizon: 64
• Action steps: 32
• Checkpoint: EMA weights plus LeRobot preprocessor and postprocessor

Evaluation

Evaluated in AlohaInsertion-v0 with 400-step episodes and one synchronous batch of 50 rollouts using seed 1000:

• Success: 3/50 (6%)
• Average reward sum: 180.88
• Average maximum reward: 1.96

Maximum-reward stage counts were 10, 10, 5, 22, and 3 for stages 0 through 4.

Notes

This is a research checkpoint trained with a LeRobot development fork. It is intended for the matching LeRobot Diffusion Policy implementation and ALOHA simulation setup. Results may vary with environment and dependency versions.