Residual-Conservative Model Predictive Path Integral Control

What the math gives to ML

The paper's transferable contribution is an uncertainty-aware control principle: when observed model residuals are large, both safety margins and the randomness of rollout-based decisions should increase, because cost rankings computed by an inaccurate predictor are unreliable. This couples residual-dependent constraint tightening with temperature relaxation rather than treating the penalty or sampling temperature as fixed hyperparameters. A direct neural-network use is to apply the same rule to world-model trajectory samplers, diffusion policies, or other candidate-action optimizers, using prediction error or ensemble disagreement as the residual signal. The key falsifiable prediction is improved failure rate and calibration under distribution shift, with only a modest loss in nominal control quality.