ARX5 seven-task ฯ0.5 fine-tune
This repository contains the last deployment checkpoint, step 9999, from a completed 10,000-step ARX5 fine-tuning run. It is not a selected-best checkpoint.
Model and training
- Initialization: official OpenPI ฯ0.5 base checkpoint,
gs://openpi-assets/checkpoints/pi05_base(upstream); not another embodiment's fine-tune or the released ARX5 baseline. - Data: all seven Table30v2 ARX5 tasks; 7,270 episodes and 8,864,335 frames.
- Inputs: three cameras and embodiment-specific robot state.
- Actions: six joint deltas plus an absolute gripper value in meters.
- Recipe: full fine-tuning, four A800 GPUs, global batch 16, FSDP 4, peak learning rate 1e-5, warmup 200 steps, EMA disabled.
- Run:
run-257947-20261027T160102, completed 2026-10-27.
Download and checkpoint layout
hf download engageassume/arx5 --local-dir ./arx5
For a private repository, authenticate with an account/token that has access before downloading. Use ./arx5 as the checkpoint root with the matching ARX5 OpenPI fine-tuning/inference configuration. Keep both components together:
params/
assets/arx5_seven_tasks/norm_stats.json
Weights use the original JAX/Orbax OCDBT format, not Transformers/PyTorch safetensors. These are deployment parameters and normalization assets, not optimizer state for exact training resume. A matching model configuration and embodiment-specific input/action transforms are required; downloading this repository alone does not install an inference runtime.
Evaluation limits
All 10,000 recorded training losses were finite. Final training loss was approximately 0.0021845654; the final 100-step mean was approximately 0.0060905743. These are training metrics, not held-out performance or real-robot success rates. No checkpoint selection or closed-loop evaluation was performed.
No new license grant is asserted here. Upstream weight and training-data licensing terms must be checked before redistribution or use.