Guanli001/grootn16-isaacsim

GR00T N1.6 — SO-101 Isaac Sim vial placement

SO-101 机械臂在 Isaac Sim 中抓起试管并放入黄色试管架的 GR00T N1.6 微调权重。

Task: Pick up the vial and place it in the yellow rack.

Checkpoint and training

• Source run: grootn16-course-sim75-plus-strict87-ovrstats-s42-lr1e5-20000-v1.
• Checkpoint: checkpoint-20000, verified global step 20,000.
• Continued from aravindhs-NV/grootn16-finetune_sreetz-so101_teleop_vials_rack_left/checkpoint-10000.
• This fine-tuning run used 87 demonstrations (51,332 frames) from so101-vials-strict-merged-v3-87-train-v21.
• Seed 42; learning rate 1e-5; dataset statistics override enabled.
• Task action horizon: 16. The model configuration retains a padded action horizon of 50 and maximum action dimension of 128; these are not the physical robot's joint count.
• The original training dataset is not included. No benchmark success rate or hardware validation claim is made by this upload.

Files and use

The two model-*.safetensors files are shards of one model. Keep both shards, the index, configuration, processor configuration, embodiment mapping, and statistics together.

Download the inference checkpoint:

hf download Guanli001/grootn16-isaacsim --local-dir ./grootn16-isaacsim

Load the downloaded directory as the model path with a compatible GR00T N1.6 runtime. This repository is a checkpoint, not a standalone robot-control application. Camera names, joint ordering, units, embodiment selection, and action processing must match the task configuration. See processor_config.json and experiment_cfg/ for the saved settings. Some original experiment configuration entries contain source-machine paths; adapt those paths when reproducing the setup.

This upload contains inference weights and configuration. Optimizer, scheduler, RNG state, and full trainer state are omitted; it is not an exact training-resume bundle.

SHA256SUMS records local SHA-256 digests of the uploaded model and supporting files. training_summary.json records the selected checkpoint's provenance.

License and attribution

This checkpoint derives from NVIDIA GR00T N1.6 and the SO-101 vial-task checkpoint by aravindhs-NV linked above. The upstream NVIDIA license is included verbatim in LICENSE and applies to this derivative. It restricts use to non-commercial research or evaluation (Section 3.3). This upload does not grant additional rights to upstream components.