Task 000759 PoseVLA, step 10,000
Inference-only SG2 PoseVLA checkpoint for Move Vegetable Whiteroom. Trained with three RGB cameras (left head, left wrist, right wrist), ordered 22-D state/action, 16-action chunks, and 224 x 224 padded images. This is the saved step-10,000 checkpoint from an intended 10,584-step run; it is not the final-step checkpoint.
The repository contains the model weights, configuration, tokenizer files, and Cyclo input metadata. Optimizer and scheduler state are intentionally excluded. The original training-machine tokenizer path in config.json was replaced with tokenizer; the Cyclo PoseVLA worker resolves it to the bundled tokenizer directory on load.
PoseVLA source revision: 1520c6b595fe75de4efe587dc5b5c2716376e4ae. LeRobot revision: 638d411cd3acf32c28d8c2120f3c41bda8bb15d4.
model.safetensors SHA-256: ff8ac3dcc2dc7c2e36dfc8aa2035035c45f9c8c7073fec1ba60a32ac345bb41a.
The safetensors structure and SHA-256 were verified on the training server. The full checkpoint loaded successfully on CPU with the pinned PoseVLA and LeRobot revisions on 2026-10-19. End-to-end action inference and robot task success have not been validated for this published bundle.
- Downloads last month
- 26