Pi05 RoboTwin

Original pi0.5 (history_action_conditioning=False). Fine-tuned on the merged RoboTwin-Clean LeRobot dataset.

Checkpoints

  • checkpoints/30000: training step 30,000 (3w)
  • checkpoints/59999: final checkpoint of the 60,000-step run (6w)

Each checkpoint contains inference parameters under params/ and normalization statistics under assets/openpi_robotwin_clean_merged/norm_stats.json. Optimizer and training state are intentionally omitted.

Training configuration

  • Base model: pi0.5
  • Action horizon: 50
  • Action dimension: 32
  • Dataset: openpi_robotwin_clean_merged
  • Seed: 42
  • Precision: bfloat16
  • Fine-tuning steps: 60,000

Use a checkpoint directory (for example checkpoints/59999) as the OpenPI checkpoint path.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading