groot-task0003-action-v6-80k

GR00T N1.7 policy trained on Task0003 SeparateRecycling to 80,000 optimizer steps.

Visual initialization: action-supervised V6 best checkpoint (vision step 12,000). The visual backbone was frozen during policy training.

Final logged training loss: 0.0069. This is a training metric, not a held-out evaluation score.

This repository contains the final policy weights, configuration, processor configuration, embodiment mapping, and normalization statistics. Optimizer, scheduler, and RNG state are retained in the original training checkpoint and are not included here.

Load with the compatible Isaac-GR00T N1.7 policy runtime. The configuration references nvidia/Cosmos-Reason2-2B for the backbone architecture and processor; access to that model and its processor assets is required. Custom Task0003 embodiment modalities are described in the included processor configuration.