REFLECT Real-world Failure-Led Embodied Continual Training

Most robot policies are trained once, deployed once, and frozen — when reality diverges from simulation, the loop stops. REFLECT treats deployment as the start of learning: an ACT policy trains in Antioch, passes a physics-based sim gate, and deploys to the real SO-101. When the arm fails, an embodied reasoning critic watches the rollout video, diagnoses what went wrong, and turns that failure into targeted simulation curriculum. Antioch generates corrective experience at scale, the policy retrains, and the improved version redeploys. One real mistake becomes the next training phase — perception, reasoning, and action closing the loop end to end.