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
Each <step>/ (10000/, 15000/, …, 45000/, 49999/ = final; there is no 5000/, it was overwritten before upload) holds params/ (orbax) and assets/ (norm stats).
The optimizer state is not included, so these checkpoints are for inference or fine-tuning, not for resuming training.
Serving
uv run scripts/serve_policy.py policy:checkpoint --policy.config pi05_atomic65_50k --policy.dir <local copy of <step>/>
Evaluation results (65 atomic tasks, split pretrain, 5 then 20 episodes per task) are added here as they come in.