Research Analyzer
← Back ICRA 2026

Whole-Body Safe Control of Robotic Systems with Koopman Neural Dynamics

Sebin Jung, Abulikemu Abuduweili, Jiaxing Li, Changliu Liu

PDF

AI summary

Key figure (auto-extracted from paper)
A unified Koopman-MPC framework learns linearized dynamics from data and embeds safety constraints directly into a single quadratic program, enabling real-time collision avoidance for complex robots.
Koopman operator safe control model predictive control neural dynamics real-time robotics safety constraints

Problem

Real-time safe control of high-dimensional, nonlinear robots is hindered by computationally prohibitive nonlinear optimization and frequent constraint infeasibility at safety boundaries.

Approach

The method learns a Koopman operator to globally linearize robot dynamics in a lifted space, then integrates a redesigned safety specification directly into a single MPC quadratic program, using adversarial fine-tuning to guarantee constraint feasibility.

Key results

  • Unified Koopman-MPC formulation solves tracking and safety in a single QP
  • Adversarial fine-tuning of the safety index drastically reduces QP infeasibility at safe set boundaries
  • Successful sim-to-real deployment on a Kinova Gen3 manipulator with minimal retraining
  • Effective obstacle avoidance and trajectory tracking demonstrated on fixed-base and floating-base robots

Why it matters

Provides a scalable, real-time safe control solution for high-DOF robots that eliminates the need for separate safety filters or exact dynamic models, accelerating practical deployment in cluttered environments.

Abstract

Controlling robots with strongly nonlinear, high- dimensional dynamics remains challenging, as direct nonlinear optimization with safety constraints is often intractable in real time. The Koopman operator offers a way to represent nonlin- ear systems linearly in a lifted space, enabling the use of efficient linear control. We propose a data-driven framework that learns a Koopman embedding and operator from data, and integrates the resulting linear model with the Safe Set Algorithm (SSA). This allows the tracking and safety constraints to be solved in a single quadratic program (QP), ensuring feasibility and optimality without a separate safety filter. We validate the method on a Kinova Gen3 manipulator and a Go2 quadruped, showing accurate tracking and obstacle avoidance.

Index terms

Robot Safety Model Learning for Control Deep Learning Methods

Related papers