robotwin_piper_x_new_sft_step20000

Pi0.5 Piper — revised dataset, checkpoint 20,000

BF16 PyTorch conversion of the OpenPI pi05_piper_new checkpoint at global training step 20,000 from experiment robotwin_piper_x_new_sft.

Training

• Initialized from gs://openpi-assets/checkpoints/pi05_base/params.
• Dataset: SidneyXie/robotwin_piper_x_20_tasks_lerobot_v21_new, revision 9f7b059913c3af6bc59625507f06692f74af15c5.
• 12,000 episodes, 2,398,063 frames; global batch size 64.
• Model: Pi0.5, action dimension 32, action horizon 50, discrete state input.
• Preprocessing: 14-dimensional state and actions, delta joint actions with absolute grippers, adapt_to_pi=False, task prompts and quantile normalization.
• Cameras: cam_high, cam_left_wrist, cam_right_wrist.
• Five videos (episodes 371, 1176, 9878, 9970, 10065) were one frame shorter than their episode metadata. One copy of the last valid image was appended to each affected video. Existing decoded frames, states and actions were preserved.

Export and inference

Converted using examples/convert_jax_model_to_pytorch.py with configuration pi05_piper_new and BF16 precision.

• model.safetensors: inference weights.
• config.json: model configuration.
• assets/robotwin_piper_x_20_tasks_lerobot_v21_new/norm_stats.json: normalization statistics copied from the source checkpoint.
• conversion_info.json: source checkpoint and conversion validation details.

Use the OpenPI PyTorch implementation with the matching pi05_piper_new configuration and the included normalization assets. This export does not include optimizer state for resuming JAX training.

Validation

Conversion input keys and exported tensor keys, shapes, BF16 dtype and finite values were checked. Normalization assets were copied exactly from the source checkpoint. JAX/PyTorch forward-pass numerical parity was not tested.