研究将采样式 MPC 统一为 Sequential Monte Carlo 推理框架
Sampling-based Model Predictive Control as Sequential Monte Carlo Inference
做机器人控制的朋友可以看看这篇：作者把 MPPI、cross-entropy method 这些常见 MPC 算法统一成一个 SMC 推理框架，只用三个维度就能配置出所有变体。重采样在长时程上把碰撞率降了 30%，push-T 实机成功率从 10% 拉到 90%，比单独调某一种算法更系统。
一篇 arXiv 论文提出，predictive sampling、MPPI、cross-entropy method 及其 biased、colored-noise、annealed 变体这类采样式模型预测控制（MPC）算法，本质上都可用一个 Sequential Monte Carlo（SMC）推理算法统一实现。该框架由三个独立维度定义：proposal、将 cost 加权粒子收缩为名义规划的 update rule、以及触发重采样的 effective-sample-size 阈值。
实验覆盖 MuJoCo 非线性控制、障碍场开环导航、7-DoF 运动规划和 non-prehensile push-T，仿真与 Franka FR3 实机均有测试。重采样能缓解长时程下的粒子退化问题：在导航域 28-knot 时程上，碰撞率比无重采样基线降低 30%。
可编程 proposal 在 push-T 任务上优势明显，仿真成功率从 1% 提升到 100%，实机从 10% 提升到 90%。这一统一视角意味着现有多种 MPC 算法只是同一推理过程在不同维度上的配置选择，便于机器人控制研究者系统性地组合与改进算法。
Sampling-based Model Predictive Control as Sequential Monte Carlo Inference
Sampling-based model predictive control (MPC) is treated as a family of distinct algorithms: predictive sampling, MPPI, the cross-entropy method, and their biased, colored-noise, and annealed variants all sample control sequences and reweight them by cost. We show a single Sequential Monte Carlo (SMC) inference algorithm implements them all, specified by three independent axes: the proposal, the update rule that reduces cost-weighted particles to a nominal plan, and the effective-sample-size threshold for resampling. We evaluate SMC-based planning on MuJoCo non-linear control, open-loop navigation through obstacle fields, 7-DoF motion planning, and non-prehensile push-T, in simulation and on a Franka FR3. Resampling combats particle degeneracy at long horizons. At a 28-knot horizon in our navigation domain it cut the collision rate by 30% over non-resampling baselines. Programmable proposals improve push-T success over perturbative baselines, improving from 1% to 100% in simulation and 10% to 90% on hardware.
This report is categorized under AITOP's "Papers" desk. Key entities and technical topics under continuous tracking include MPC, Sequential Monte Carlo, MPPI, 机器人控制.
Source attribution: curated from arXiv cs.LG (@Austin Garrett, Muyang Yan, Zachary Kingston).