Research Analyzer
← Back ICRA 2026

SIT-LMPC: Safe Information-Theoretic Learning Model Predictive Control for Iterative Tasks

Zirui Zang, Ahmad Amine, Nick-Marios Kokolakis, Truong Xuan Nghiem, Ugo Rosolia, Rahul Mangharam

PDF

AI summary

Key figure (auto-extracted from paper)
SIT-LMPC enables robots to iteratively improve performance in uncertain environments while strictly satisfying safety constraints through parallelized information-theoretic control and learned uncertainty models.
Safe MPC Iterative Learning Control Information-Theoretic Control Normalizing Flows GPU Parallelization Stochastic Control

Problem

Iterative robotic tasks in uncertain environments require balancing safety, robustness, and performance, but existing constrained model predictive control methods are often overly conservative, lack general constraint handling, or cannot run in real-time.

Approach

The authors introduce SIT-LMPC, which combines information-theoretic MPC with an online adaptive penalty method to enforce constraints dynamically, while using normalizing flows to learn complex cost distributions from past trajectories, all optimized for GPU parallelization.

Key results

  • Online adaptive penalty method dynamically balances safety and optimality
  • Normalizing flows model iteration costs with richer uncertainty than Gaussian priors
  • Achieves 100Hz+ real-time control on hardware via full GPU parallelization
  • Demonstrates iterative performance gains and robust constraint satisfaction in simulations and experiments

Why it matters

Provides a scalable, real-time control framework for safe iterative robotics, advancing applications in autonomous navigation, aerial systems, and precision automation.

Abstract

Robots executing iterative tasks in complex, un- certain environments require control strategies that balance robustness, safety, and high performance. This paper introduces a safe information-theoretic learning model predictive control (SIT- LMPC) algorithm for iterative tasks. Specifically, we design an iterative control framework based on an information-theoretic model predictive control algorithm to address a constrained infinite-horizon optimal control problem for discrete-time nonlin- ear stochastic systems. An adaptive penalty method is developed to ensure safety while balancing optimality. Trajectories from previous iterations are utilized to learn a value function using normalizing flows, which enables richer uncertainty modeling compared to Gaussian priors. SIT-LMPC is designed for highly parallel execution on graphics processing units, allowing efficient real-time optimization. Benchmark simulations and hardware experiments demonstrate that SIT-LMPC iteratively improves system performance while robustly satisfying system constraints.

Index terms

Optimization and Optimal Control Machine Learning for Robot Control Planning under Uncertainty

Related papers