AIRO-Doffy-DP-vision-joint-matched-20260914

Matched RGB-and-joint original_dp Diffusion Policy baseline.

• Inputs: RGB observations and seven normalized joint values.
• Visual conditioning: enabled for two observation steps; Beaver conditioning is disabled.
• Native DDPM settings: horizon 16, action steps 8, and 100 diffusion timesteps.
• Execution gate: disabled.
• Dataset: WRM_grasp_cylinder_different_sizes_lero_recollect_gray_tightness, with an explicit 100-episode training split and 25-episode validation split (episodes 20--24, 45--49, 70--74, 95--99, and 120--124).
• Training: seed 42, batch size 32, six DataLoader workers, and final step 100000; W&B online tracking is required.
• EMA weights are retained in last.pt for deployment.

The checkpoints/ directory contains the numbered original_dp_step_*.pt milestones, last.pt, the resolved configuration, metrics, and the source and dataset checksum artifacts. W&B directories and credentials are excluded.

SHA-256 of checkpoints/last.pt: 3ff1bf9683e38c280234b39810e0df0f6aa614a082f74e367e0d1b02ff914d1e.