MPPI's exploration covariance is also its feedback gain: too little exploration destabilizes an inverted pendulum
Synopsis
In model predictive path integral (MPPI) control with a warm start, the exploration covariance sets not only how widely candidates are sampled but also the closed-loop feedback gain. For linear quadratic problems with infinitely many samples, too little exploration makes MPPI nearly repeat the previous solution and fail to stabilize an exponentially unstable plant, while the same covariance lowers the effective sample size, a structural conflict between stability and sample efficiency. Stabilizing Subspace MPPI (SS-MPPI) computes the directions needed for stabilization and samples only the rest, staying stable with small exploration and few samples in inverted-pendulum and quadruped simulations.
Interpretation
In MPPI with a warm start, the exploration covariance also acts as the closed-loop feedback gain, so insufficient exploration can destabilize the loop rather than merely bias the solution toward the previous control sequence. Earlier stability analyses treated a small exploration covariance only as a bias toward the warm-started solution and required low temperature and enough samples; whether too little exploration destabilizes the closed loop had remained open. For linear quadratic problems in the limit of infinitely many samples, the MPPI update is an interpolation between the warm-started nominal sequence and the exact MPC solution, with the interpolation weight set by the exploration covariance; sweeping the exploration variance on the linearized inverted pendulum, MPPI is stable only above a threshold, and the computed spectral radius and effective-sample-size bound match Monte Carlo to within one percent.
Large exploration for stability and small exploration for sample efficiency are carried by one covariance, so the conflict is structural and neither more samples nor temperature tuning resolves it. Earlier work treated the exploration covariance as a free hyperparameter of a single update and studied update quality and sample requirements separately, without placing the across-time feedback role and inference efficiency in one framework. For linear quadratic problems, an upper bound on the normalized effective sample size is expressed through the feedback gain, and temperature enters the closed loop and the bound only as a scale of the covariance; on the inverted pendulum the effective sample size of MPPI falls below one half at the threshold, while covariance-shaping CoVO-MPC moves the threshold about tenfold.
SS-MPPI computes the directions of the control sequence needed for stabilization and samples the rest, keeping the closed loop stable at small exploration with the effective sample size close to one. Earlier sampling-plus-computation schemes let the designer, the cost or the model choose which part to compute; SS-MPPI chooses the computed subspace by the closed-loop stability of the plant and the warm start. On the inverted pendulum, computing only the first input keeps SS-MPPI stable across the sweep with effective sample size above one half; for a Unitree Go2 walking forward at 1 m/s in Genesis, the standing-configuration linearization selects the first two input blocks, and SS-MPPI walks at every tested variance and keeps roughly its distance when samples are reduced.
Perspective
The rigorous part concerns linear quadratic problems in the limit of infinitely many samples; the nonlinear case gives a proposition on an exponentially attracting fixed point near an equilibrium, and the quadruped walking at 1 m/s is a Genesis simulation whose stabilizing subspace and terminal weight come from the standing-configuration linearization.
The nonlinear stability guarantee holds only near an equilibrium, and walking has no fixed equilibrium, so the quadruped results are empirical validation in simulation; with a nonquadratic cost the weighted sample mean need not be the minimizer, so the fixed point may differ from the origin.
