Research Analyzer
← Back ICRA 2026

SafeFlowMPC: Predictive and Safe Trajectory Planning for Robot Manipulators with Learning-Based Policies

Thies Oelerich, Gerald Ebmer, Christian Hartl-Nesic, Andreas Kugi

PDF

AI summary

Key figure (auto-extracted from paper)
SafeFlowMPC combines flow-matching learning with real-time optimization to guarantee hard safety constraints while maintaining the flexibility and generalization of learning-based trajectory planning.
Safe trajectory planning Flow matching Model-predictive control Robot manipulators Learning-based control Real-time optimization

Problem

Learning-based trajectory planners lack rigorous safety guarantees, while optimization-based methods struggle with flexibility and real-time reactivity in dynamic environments.

Approach

The method integrates a flow-matching model that learns desired behaviors from demonstrations with a suboptimal model-predictive control backend that continuously projects trajectories onto a safety manifold to enforce hard constraints in real time.

Key results

  • Novel safe flow-matching procedure integrating learning with real-time optimization
  • Guaranteed safety at all times via enforced safe terminal constraints
  • Successful reactive motion planning demonstrated on a real-world 7-DoF KUKA manipulator
  • Strong performance in dynamic grasping and human-robot object handover tasks

Why it matters

Provides a scalable, real-time solution for safe robot manipulation in dynamic environments, bridging the gap between data-driven adaptability and classical control safety.

Abstract

The emerging integration of robots into everyday life brings several major challenges. Compared to classical industrial applications, more flexibility is needed in combination with real-time reactivity. Learning-based methods can train powerful policies based on demonstrated trajectories, such that the robot generalizes a task to similar situations. However, these black-box models lack interpretability and rigorous safety guar- antees. Optimization-based methods provide these guarantees but lack the required flexibility and generalization capabilities. This work proposes SafeFlowMPC, a combination of flow matching and online optimization to combine the strengths of learning and optimization. This method guarantees safety at all times and is designed to meet the demands of real-time execution by using a suboptimal model-predictive control for- mulation. SafeFlowMPC achieves strong performance in three real-world experiments on a KUKA 7-DoF manipulator, namely two grasping experiment and a dynamic human-robot object handover experiment. A video of the experiments is available at https://www.acin.tuwien.ac.at/en/42d6. The code is available at https://github.com/TU-Wien-ACIN-CDS/SafeFlowMPC.

Index terms

Deep Learning in Grasping and Manipulation Robot Safety Learning from Demonstration

Related papers