AEGIS: A Backup Reflex for Physical AI

In three linesAEGIS detects high-risk steps in long-horizon robot manipulation by probing frozen activations of a weak policy. Upon detection, control switches to a stronger policy for only those steps. On LIBERO-Spatial, AEGIS recovers 10.1% of lost trajectories (vs 4.6% for blind escalation), activating the stronger policy on only 38% of steps.

Summary generated by Claude — human-verified