ducido/pi05_atomic65_50k

π0.5 trained on the 65 RoboCasa365 atomic tasks: one 50k-step cosine schedule

openpi (JAX) π0.5 checkpoint, config pi05_atomic65_50k of the openpi RoboCasa fork. It is the clean, single-schedule version of ducido/pi05_atomic65_from_base, whose 50k lineage restarted the optimizer three times (10k → 15k → 50k).

Init	pi05_base (openpi)
Data	RoboCasa365 pretrain split, the 65 atomic tasks (all_atomic_tasks), human demos only, 7356 episodes / 1.50M frames; no composite-task data
Schedule	50k steps, batch 64, warmup 1000, cosine 2.5e-5 → 2.5e-6 (~2.1 epochs)
Hardware	4×A100-80GB, FSDP 4

Layout

Two step folders are kept, each with params/ (orbax) and assets/ (norm stats):

• 35000/: the best checkpoint of this run on the evaluation below;
• 49999/: the end of the schedule.

The other step folders (10000/ … 40000/) were deleted on 2026-10-22, following a decision to keep only the best checkpoint. The optimizer state is not included, so the checkpoint is for inference or fine-tuning, not for resuming training.

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

The client is openpi eval_g0, replanning every 5 steps, with seeds 700000 + episode for episodes 0–4 (325 episodes per checkpoint). The table gives the mean success rate over the 65 tasks.

Step	10k	15k	20k	25k	30k	35k	40k	50k (49999)
SR	57.2	60.0	64.0	65.2	66.2	70.5	68.3	69.2

• Reference: RoboCasa365 pi05_pretrain_human300 75k, which was trained on 300 tasks, scores 40.3% on the same episodes (43.7% with 20 episodes per task).
• Noise: with 5 episodes per task, the 95% CI is about ±5 points. The curve is flat from 35k: 35k vs 40k p = 0.42 and 35k vs 49999 p = 0.69 (paired McNemar), while 35k vs 25k is significant (p = 0.043).

Serving

uv run scripts/serve_policy.py policy:checkpoint --policy.config pi05_atomic65_50k --policy.dir <local copy of 35000/>