groot-task0003-v4-40k

GR00T N1.7 policy checkpoint at 40,000 cumulative optimizer steps on Task0003 SeparateRecycling.

Visual initialization: V4 vision-tuned encoder. The visual backbone was frozen during policy training.

Training loss logged at step 40,000: 0.0135. This is a training metric, not a held-out evaluation result.

Contains the original 40K policy weight shards and index, model and processor configuration, embodiment mapping, and normalization statistics. Optimizer, scheduler, RNG state, and training logs are excluded.

Use a compatible Isaac-GR00T N1.7 policy runtime. Configuration references nvidia/Cosmos-Reason2-2B for the backbone architecture and processor assets, which must be accessible. Task0003 modality definitions are included in processor_config.json.