pi05-piperx-full558-intervention-value-1to1-step35999

Pi05 PiperX: original demonstrations + intervention/value segments

Final FP32 policy after 35,999 additional training updates from the original full558 50,000-step checkpoint. Training finished on 2026-09-10 (Asia/Shanghai).

This export contains model weights, configuration, normalization processors and tokenizer files. Optimizer, scheduler and RNG states are not included. Optimizer hyperparameters in the JSON configuration are descriptive settings, not optimizer state.

Training

• Eight GPUs; batch size 8 per GPU, global batch size 64; seed 1000.
• Each global batch contains 32 original demonstration frames and 32 frames from the combined intervention/value dataset.
• Original dataset: 558 episodes, 1,255,729 frames.
• Combined dataset: 7,762 episodes, 460,775 frames (138,514 human intervention frames and 322,261 value-selected frames). Sampling within this dataset is uniform over frames.
• Value-selected segments are selected demonstration fragments. This run is supervised policy training, not an RL rollout update.
• Action horizon 50; 14 physical action dimensions; padded action timesteps are excluded from the training loss (mask_action_padding_loss=true).
• Learning rate: 200-update warmup, peak 2.5e-6, cosine decay to 2.5e-7.
• Last 1,000 updates: mean training loss 0.00455406, averaged over all eight ranks.

The loss plot shows all 35,999 updates, trailing 100/500-update averages, and the recorded learning rate. Training loss is not a measure of real-robot task success.

Runtime and compatibility

Training used Python 3.12.3, PyTorch 2.10.0+cu128, Transformers 5.3.0 and LeRobot 0.6.1 (upstream commit 7e241bd630a3719a56157a497ce5d08f244784f1, with local intervention sampling and action-padding-mask patches).

Inputs are three RGB cameras (base_0_rgb, left_wrist_0_rgb, right_wrist_0_rgb) and a 14-dimensional state. Outputs are 50-step chunks of 14-dimensional actions. Use the saved preprocessing/postprocessing and normalization tensors.

The JSON files preserve training-time paths for provenance. When loading on another machine, pass the downloaded model directory as the pretrained path and override the tokenizer processor path to its local tokenizer subdirectory. For further training, the compatible LeRobot implementation must support the intervention sampling and action-padding-mask fields; do not silently discard padding masks.

Validation

Before switching the training machine to GPU-free mode, the final checkpoint was reloaded for a real-data forward check and action decoding. The decoded action tensor had shape [2, 50, 14] and contained only finite values. All eight training logs contain the complete 1–35,999 update sequence with finite loss, gradient norms and learning rates. Padding-mask tests covered short fragments and verified that padded actions do not affect loss or gradients.

model-export-manifest.json lists the exported model files with byte sizes and SHA-256 digests. No real-robot success-rate evaluation is included in this export.