SmolVLA-RLT: PegInsertionSide-v1 (three cameras)

Successful motion-planning demonstrations replayed from saved ManiSkill states. Every 20 FPS observation provides 512×512 RGB from environment_camera, hand_camera, and insertion_camera. The insertion camera is mounted in the peg frame with its reference at the red insertion end. Actions are 8D absolute pd_joint_pos commands; observation.state is 9D Panda qpos. Exact source and camera-config hashes are in render_manifest.json.