Geoffrey19/cei2-pi05-mix5-112h

CEI-2.0 pi05 - 5-embodiment mixture (112 container-hours)

A pi05 policy fine-tuned on a 5-embodiment mixture dataset built with CEI-2.0, a rendering-grounded cross-embodiment augmentation pipeline that re-paints DROID demonstrations into other robot embodiments.

What this checkpoint is for

It is one arm of a controlled comparison at a fixed data budget. Every run in the set trains on the same number of frames with the same recipe, so differences between them reflect the variable under test rather than data volume:

run	variable	frames
ur7e_xhand 112h	single embodiment, RoboGS render (reference)	12,089,174
mix5 112h	5 embodiments, 22.4 h each	12,099,508
mix10 112h	10 embodiments, 11.2 h each	12,100,916
droid112 112h	the same episodes, real unrendered DROID footage	12,089,174

This repository is mix5: the 112 h budget split evenly across five embodiments (22.4 h each), trained as five separate datasets so each keeps its own joint dimension and its own normalisation statistics.

• embodiments: ur7e_xhand, ur7e_leaphand, fr3v2_xhand, fr3v2_leaphand, ur7e_ruihand
• episodes / frames: 33,920 / 12,099,508
• normalisation statistics included for: fr3v2_leaphand, fr3v2_xhand, ur7e_leaphand, ur7e_ruihand, ur7e_xhand

Training

Identical to the reference run, so the comparison isolates the variable above:

• base model pi05_base, action dim 32, action horizon 16
• 50,000 steps, global batch 1024, 4 nodes x 8 GPUs (DDP)
• lr 5e-5 constant after a 1,000-step warmup, AdamW, grad clip 1.0, EMA 0.999
• one exterior camera per episode; prompts are DROID's own language annotations
• state and actions are the embodiment's absolute qpos

Files

• model.safetensors - the policy weights (812 tensors)
• metadata.pt - openpi training metadata
• assets/<embodiment>/norm_stats.json - normalisation statistics, required at inference: a policy run without the statistics it was trained under will produce systematically wrong actions
• the optimizer state (12.8 GB) is not included; it is only needed to resume training

Hours convention

"112h" counts container hours (frames / 30 fps). The underlying DROID recordings run at about 14.1 Hz, so the real robot time is roughly 2.1x larger (~238 h). Published figures that count wall-clock hours are not directly comparable.

Caveats

• Held-out evaluation uses 20 trajectories (640 samples); the observed noise floor is about 0.06 deg, so small differences between runs are not meaningful.
• The mixtures carry no embodiment identifier in the prompt by design: the policy must infer which body it is driving from the image.
• Per-embodiment error figures are in different joint spaces (18-27 dims depending on the hand, 8 for real DROID), so degree errors are not directly comparable across arms of the comparison.

License

PolyForm Noncommercial 1.0.0 - noncommercial research use only. DROID, and the model and rendering components used to build the training data, carry their own licenses.