Clustering-Embedded MPPI: Avoiding Averaging-Induced Failure and Enabling Efficient Cluster Selection for Dynamic Obstacles

Zidong Liu, Kaixin Chang, Xu Chen

Department of Mechanical Engineering, University of Washington

Published and presented at IROS 2026

Supported by

UW + Amazon Science Hub

arXiv Video BibTeX

Real-world 6-DoF UR5e obstacle-avoidance experiment using CE-MPPI. Default playback: 2.5x.

Abstract

With the widespread availability of parallel computing hardware, sampling-based motion planning methods such as Model Predictive Path Integral (MPPI) control have become increasingly powerful for complex nonlinear systems in non-smooth task spaces. However, MPPI can suffer from averaging-induced failure in cluttered environments, where the importance-weighted update averages incompatible rollouts and leads to hesitation or collision.

We propose Clustering-Embedded MPPI (CE-MPPI), a framework that structurally resolves this failure mode by pruning colliding trajectories, clustering feasible rollouts with a collision-derived geometric direction feature, and selecting a single obstacle-aware cluster for the control update. In dynamic scenes, CE-MPPI selects clusters that counter obstacle motion, reducing persistent coupling with moving obstacles.

Method Overview

CE-MPPI samples perturbed control sequences and performs parallel forward simulation to obtain trajectory rollouts and costs. Colliding rollouts are pruned, the remaining feasible rollouts are clustered via DBSCAN, and a single cluster is selected (static: minimum average cost; dynamic: motion opposite to the obstacle direction). The MPPI update is then computed using weights normalized over the selected cluster.

Experimental Results

48%

reduction in real-world time-to-goal versus standard MPPI

12%

shorter real-world end-effector path

17.4%

reduction in dynamic-scenario time-to-goal versus CSC-MPPI

9.1%

shorter simulated path length versus CSC-MPPI

Simulation and Real-World Demonstrations

2-D JAX Simulation

Dynamic-obstacle replanning with clustering-based mode selection. Default playback: 2x.

Real-World UR5e Experiment

High-dimensional obstacle avoidance with CUDA-parallel rollouts in Isaac Gym. Default playback: 2.5x.

Paper

The submitted manuscript is embedded below. If the preview does not load, use the direct paper link in the header.

BibTeX

@inproceedings{liu2026cemppi,
  title = {Clustering-Embedded MPPI:
           Avoiding Averaging-Induced Failure and Enabling Efficient
           Cluster Selection for Dynamic Obstacles},
  author = {Liu, Zidong and Chang, Kaixin and Chen, Xu},
  booktitle = {2026 IEEE/RSJ International Conference on
               Intelligent Robots and Systems (IROS)},
  year = {2026},
  eprint = {2607.06499},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2607.06499}
}