Skill UI

Provides the workflow to run NVIDIA Cosmos Policy evaluations on the LIBERO and RoboCasa benchmarks, covering headless EGL rendering, smoke and full runs, and inference latency profiling for robot manipulation scenarios.

Fine-tunes and evaluates OpenVLA-OFT/OFT+ to generate robot actions via LoRA-adapted VLA backbones, continuous L1 or diffusion action heads, and FiLM conditioning, covering LIBERO simulation and ALOHA real-world deployments with official reproduction workflows.

Fine-tune and serve Physical Intelligence OpenPI pi0 family models with JAX or PyTorch backends for robot manipulation across ALOHA, DROID, and LIBERO; includes workspace setup, checkpoint conversion, norm stat prep, and WebSocket policy inference for debugging and deployment.