Skip to main content
Back

UT Austin's UGE-MPC Beats MPPI With 88% Success Rate in Cluttered Robot Navigation

A new sampling scheme for model predictive control spreads candidate trajectories apart using uncertainty ellipsoids, cutting convergence time by up to 72 percent.

Pro

Paper, Full Access
  • UGE-MPC models each rollout as a Gaussian trajectory distribution using propagated uncertainty ellipsoids.
  • Hellinger distance between distributions forces candidates apart in configuration space, not just action space.
  • Achieves 88% success in cluttered navigation vs 54-64% for MPPI, MoG-MPPI, and SV-MPC baselines.
  • Reports 72.1% faster convergence in open environments and 66% faster in cluttered ones.
  • Validated on a 1/5-scale truck with lidar; MPPI failed every trial, UGE-MPC succeeded consistently.
  • Code and project details at ogpoyrazoglu.github.io/cuniform_sampling; accepted to ICRA 2026.

Sampling-based model predictive control has become the go-to method for robots that plan under messy, non-smooth cost functions: autonomous ground vehicles weaving through obstacles, or manipulators handling contact. The dominant algorithm in that family, Model Predictive Path Integral (MPPI) control, samples noisy variations of a nominal control sequence and averages them by cost. It works, but has a well-known Achilles heel. A new paper from the Robotic Sensor Network Lab at UT Austin and the University of Minnesota, Uncertainty Guided Exploratory Trajectory Optimization (UGE-TO), targets that weakness directly.

Why MPPI Gets Stuck

MPPI samples action sequences from a Gaussian distribution centered on the nominal trajectory, producing rollouts concentrated around it. When the initial guess points the wrong way, or when a cluttered environment demands a maneuver that looks nothing like the nominal path, those samples cluster in the wrong region of the state space and the controller collapses into a local minimum. Prior fixes such as Log-MPPI, entropy-regularized policies, and Stein Variational MPC add diversity by perturbing in the action space, which can miss the way system dynamics stretch and warp those perturbations over a long horizon.

Comparison of first-iteration trajectory samples for MPPI, SV-MPC, and UGE-MPC when the goal is behind the robot

Rollouts as Distributions, Not Curves

The core idea of UGE-TO is to stop treating each rollout as a single curve. Each candidate trajectory is represented as a probability distribution: the mean state sequence plus an uncertainty ellipsoid propagated forward through the linearized dynamics, capturing the combined effect of process noise and action perturbations. With every rollout now living as a Gaussian over the state trajectory, the algorithm can measure how much two candidates actually overlap in configuration space, rather than just how much their control sequences differ.