physical-ai-context git:20260703.774344bA

Immutable. This exact content is served forever at /api/v1/blob/a3429a5f2be7761d.

---
name: physical-ai-context
description: Use for Claude Code reviews that need robotics, simulation, GPU-routing, sim-to-real, or BDD100K pipeline domain context.
---

# Physical AI Context

Sim-to-real pipeline:

1. RL teacher policy in Genesis.
2. Visual demos.
3. SimToLeRobot adapter.
4. Student policy in LeRobot.
5. Evaluation in Genesis.

Isaac Lab requires RT cores: L40S or RTX Pro 6000. H100 and H200 do not have RT cores.

Genesis serverless RL teacher training works. Visual demo generation is blocked on EGL/DRI device access in containers.

LeRobot is Tier 1 validated on B300. SONIC, GR00T, Isaac Lab, and Cosmos are vendor-paced on NVIDIA CUDA 13 alignment.

Route SONIC to H100; L40S on-demand capacity is effectively zero for the required preset.

The BDD100K demo pipeline is the reference end-to-end workflow. `npa/workflows/workbench/skypilot/bdd100k-pipeline.yaml` is the canonical YAML.

Key claim: one YAML file describes a full Physical AI pipeline; SkyPilot orchestrates it on Nebius; results complete in about 30 minutes on a single H100.