Compiling Calibration into Frozen Robot Policies: An Autoresearch Study of Weight-Space Repair

Abstract

When a trained robot policy encounters a changed sensor or actuator interface, its weights can encode a competent behavior and still produce the wrong physical commands. We study whether the interface correction can be compiled into the boundary layers of a frozen neural policy using paired offline calibration. For invertible affine observations and signed-permutation actuator gains, the edit implements the same raw command as an explicit adapter. Under command limits, the true-map edit yields the closest reachable action to the nominal one at each state. In three simulated MuJoCo tasks, structured calibration with eight sensor pairs and axis-wise motor probes recovers near-clean returns in the exact condition. With noisy calibration it recovers near-clean returns on Reacher and Hopper and partially restores double-pendulum performance. In noisy Hopper, eight-pair structured and 64-pair dense calibration both attain about 0.011 action RMSE on random-action probes. In an exploratory check on nominal policy trajectories their errors are 0.017 and 0.458, respectively. Under cross-talk, dense calibration with 64 pairs gives lower open-loop error than the structured fit at the same budget. Reduced torque leaves Hopper far below its nominal return even with true calibration.

Keywords