This Diffusion Policy was trained with observation.state zeroed after normalization.
Zero it the same way at inference (robot/run_policy.py does this automatically).