masondx/decoupled-diffusion-aloha-insertion-local-state-20k-bs32

Decoupled Bimanual Diffusion — ALOHA Insertion — Local State

EMA inference checkpoint for a decoupled bimanual diffusion policy trained on lerobot/aloha_sim_insertion_human.

Configuration

• Training steps: 20,000
• Batch size: 32
• Training seed: 1000
• Communication mode: none
• State routing: split (each arm receives its own 7-dimensional state)
• Actions: two 7-dimensional arm actions
• Shared 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 synchronous batches of 50 rollouts. Across seeds 1000 and 2000 (100 episodes total):

• Success: 11/100 (11%)
• Average reward sum: 175.28
• Average maximum reward: 2.00

Per-seed success was 5/50 (seed 1000) and 6/50 (seed 2000).

Notes

This is a research checkpoint from a custom LeRobot fork implementing decoupled bimanual diffusion. It is intended for the matching policy implementation and ALOHA simulation setup. Results may vary with environment and dependency versions.