pi0.5 trained on the full RoboPRO LeRobot set (roboreal_lerobot, 15,999 episodes / 3.74M frames) with RoboResearch's pi05_base. This is a completed reproduction of the jax_30000 checkpoint from a different codebase, and it lands within 6% of it.

Both checkpoints scored on the same 3,072 samples, same config, same data slice, same sampling seeds, in raw action units — every normalization undone, which is what makes one number mean the same thing for both.

The 6.0% gap is not uniform: this run is better on slice A and worse on slice B, so the two models differ in where they are strong rather than one dominating. The reference numbers reproduce across independent measurements to five decimals, so the difference is signal rather than measurement noise.

Both models trained on this data, so this measures fit, not generalization.

Normalization statistics are the same file the original jax_30000 run used (md5 9d56ef7be163a56fcba0283ebe4a04db) and ship in assets/roboreal_lerobot/.

The run was killed at step 12,700 with no Python traceback — cause unconfirmed, dmesg was not readable to check the kernel OOM killer — and resumed from its step-12,000 checkpoint. Training metrics are continuous across the resume.

These are single eval batches, so they are noisy and not comparable to the 3,072-sample table above.

params/                       12 GB, the weights
assets/roboreal_lerobot/      norm_stats.json
_CHECKPOINT_METADATA          orbax metadata

train_state/ (31 GB optimizer state) is not uploaded — this is for inference, not resuming.

Closed-loop evaluation in RoboPRO: see docs/evaluation.md. An earlier checkpoint of this run is at robopro_pi05_meta_15000.