pi05_atomic65_from_base

Checkpoints removed (2026-10-22). Training stopped at about 47k steps, and all step folders were deleted. The best atomic-only checkpoint from this work is ducido/pi05_atomic65_50k/35000/ (70.5%). It was trained with one cosine schedule, without the optimizer restarts of this run.

A π0.5 checkpoint (openpi, JAX) fine-tuned from pi05_base on RoboCasa365 atomic tasks only. It contains no composite-task data. This is a research checkpoint.

Evaluation (RoboCasa365, 65 atomic tasks, pretrain split, 5 episodes per task)

openpi eval_g0 client, replan every 5 steps, seeds 700000 + episode, episodes 0–4. Mean success rate over the 65 tasks:

Checkpoint	10k	15k	20k	25k	30k	35k	40k	45k
Mean SR	52.0	62.5	57.5	60.6	64.6	65.2	65.5	66.5

The reference RoboCasa365 pi05_pretrain_human300 75k checkpoint, trained on 300 tasks, gets 40.3%.

With 5 episodes per task the per-task numbers are noisy; the paired differences (same seeds) are significant (McNemar p < 0.001 for 15k vs 10k and for 10k vs 75k).

openpi config	pi05_atomic65_from_base, a copy of pi05_pretrain_human300 (openpi RoboCasa fork) with only the data changed
Model	Pi0Config(pi05=True, max_token_len=200)
Init	gs://openpi-assets/checkpoints/pi05_base/params
Data	65 RoboCasa365 atomic tasks: get_ds_soup(split="pretrain", task_set="all_atomic_tasks", source="human"), i.e. human demos, pretrain split, filter key 100_demos (~100 demos per task)
Norm stats	combined from each dataset's meta/stats.json (assets/norm_stats.json)
Batch	64 (2×A100 80GB, FSDP 2)
LR (steps 0–10k, config pi05_atomic65_from_base)	cosine: warmup 1k, peak 2.5e-5, decay over 100k steps to 2.5e-6. The run stops at 10k steps, so the LR is still near peak (≈2.49e-5 at 5k, ≈2.45e-5 at 10k)
LR (10k–15k, config pi05_atomic65_cont10k)	init from 9999/ params, fresh optimizer; warmup 500, cosine planned over 20k steps from 2.5e-5 to 2.5e-6, stopped after 5k steps (LR ≈2.2e-5)
LR (15k–50k, config pi05_atomic65_cont15k)	init from 15000/ params, fresh optimizer; warmup 500, cosine over 35k steps from 2.5e-5 to 2.5e-6; stopped at step ~32k (≈47k total) on 2026-10-22

Layout

No checkpoints remain. The eval numbers above were measured before the folders were deleted.

License

These weights are derived from openpi's pi05_base (Physical Intelligence), which builds on PaliGemma. Use is subject to the license terms of those upstream weights, including the Gemma Terms of Use.