hqfang/groot-so100_101

GR00T N1.7 SO100/101

Final inference checkpoint from full GR00T N1.7-3B fine-tuning for 150,000 optimizer steps with global batch 256 on 1,209 SO100/101 LeRobot v3 datasets (19,227,195 frames). Training completed on September 23, 2026.

Training used FP32 master parameters, gradients and Adam moments with BF16 compute. Final training batch loss was 0.00721034; this is not an evaluation score. No robot rollout evaluation is reported here.

Contents

Model weights, policy configuration, pre/postprocessing configuration and normalization tensors, and VLM tokenizer/processor assets are included. Optimizer state, RNG state, datasets, logs and credentials are excluded. Weights are unchanged from checkpoint 150000. The exported configuration replaces the training machine's base-model path with nvidia/GR00T-N1.7-3B.

Compatibility and inference inputs

Use the updated LeRobot GR00T N1.7 implementation from this training project. The earlier pi05 inference code ZIP is not a drop-in supported runtime for this checkpoint: its GR00T preprocessing predates camera-padding removal and bundled VLM-processor serialization. Required processor implementation SHA256: cf0b69e95280414b1fea5ff8caac2f58ff9de949318d833fc06f856e774339a8. The training environment used Python 3.12, PyTorch 2.11 and Transformers 5.5.4.

This is a LeRobot policy checkpoint, not a raw Isaac-GR00T checkpoint. Load the saved policy AND its saved pre/postprocessors. The current policy loader also initializes from nvidia/GR00T-N1.7-3B before loading these fine-tuned weights; access to that upstream model is therefore required. For reproducibility, the base revision used in training was 2fc962b973bccdd5d8ce4f67cc63b264d6886495.

Provide a task instruction, six-dimensional robot state, and 1–4 real RGB views. Map custom camera inputs to observation.images.camera_0 through camera_3. Camera order was randomized per episode during training. If supplying padded camera slots, mark them with the corresponding <camera_key>_is_pad boolean mask so the updated processor removes them before VLM tokenization. Use a consistent order within an inference episode and set policy/processors to evaluation mode. Saved processors apply shared pooled q01/q99 normalization without clipping and convert model output back to the six action channels. The action chunk horizon is 40. Preserve the training joint ordering and units.

See the upstream model repository for applicable base-model terms.