Long horizons stop draining sampling-based control: Laguerre-MPPI decouples search dimension from horizon length
Synopsis
Sampling-based model predictive control MPPI degrades as the prediction horizon lengthens, because its search dimension grows linearly with the horizon and samples become sparse. Laguerre-MPPI represents the whole input sequence with discrete-time Laguerre basis functions and samples only their coefficients, so the search dimension is independent of the horizon. In mobile-robot obstacle avoidance the relative input variation was 0.837 for Laguerre-MPPI versus 1.354 for conventional MPPI; in step responses, extending the horizon from 5 to 35 degraded conventional MPPI substantially while Laguerre-MPPI degraded only slightly.
Interpretation
The search dimension of sampling-based model predictive control MPPI no longer grows with the prediction horizon, and maintaining sampling density over a long horizon no longer requires more samples. Conventional MPPI samples the entire input sequence directly, so the search dimension grows linearly with the horizon; with a limited sample count, extending the horizon sparsifies the samples in the search space, degrading control performance and leaving high-frequency components that cause chattering and actuator wear. In a mobile-robot obstacle-avoidance experiment the arrival times were close, 3.93 s for MPPI and 3.89 s for Laguerre-MPPI; over the interval where both follow nearly the same course, the relative input variation averaged over time and 50 random seeds was 1.354 for MPPI and 0.837 for Laguerre-MPPI, about 38% smaller.
The whole input sequence is generated as a linear combination of a few Laguerre basis-function coefficients, so sampling happens in coefficient space rather than in per-timestep input space. Discrete-time Laguerre basis functions form an orthonormal basis; each has low-pass character, and higher orders peak later, so low-order functions concentrate on the early part of the horizon while the later part is represented more coarsely. In a one-dimensional double-integrator step response with a fixed sample count and 4 basis functions, the horizon varied from 5 to 35; at shorter horizons the two methods performed comparably and improved with horizon, while at longer horizons conventional MPPI degraded markedly with larger seed-to-seed spread and Laguerre-MPPI degraded only slightly.
Perspective
The stability condition and effective-sample-size analysis are derived for linear time-invariant systems with quadratic cost, and the numerical experiments cover two-dimensional double-integrator mobile-robot obstacle avoidance, a one-dimensional double-integrator step response, and closed-loop oscillator simulations; the reference coefficient is set to zero and the coefficient-space covariance takes a specific form satisfying the consistency condition.
Stability depends on the basis selection: of 10 basis sets tested on the oscillator, 7 gave spectral radius below 1 and 3 gave above 1, predicting oscillatory divergence. The effective-sample-size comparison concerns the maximum over the initial estimate coefficient and does not guarantee the ordering for every initial estimate coefficient; the sample-count lower bound is only a necessary condition for keeping root-mean-square error within tolerance, and the error may still exceed the tolerance at some times even when the bound is met.
