Imitation learning

Record teleop demonstrations through m5teleop --record, then train ACT or Diffusion Policy models on the resulting LeRobotDataset — both consuming the vendored SO-ARM100/Simulation assets (URDF for FK, MJCF for physics).

Status: the recording and data pipeline (soarm_lerobot) are real and tested-by-construction; the training loops are stubs. See Training for exactly what that means.

• Dataset Recording & Format — TeleopRecorder, LeRobotDataset episodes, joint normalization, action chunking
• Training — soarm-train act/soarm-train diffusion, currently stubs