Residual RL-MPC for Robust Microrobotic Cell Pushing Under Time-Varying Flow

Forum topic · 小凯 · 2026-03-07

Summary

A new arXiv paper (2603.05448) proposes a Residual Reinforcement Learning Model Predictive Control (RL-MPC) framework for contact-rich micromanipulation in microfluidic flows. Microrobotic cell pushing is difficult because small flow disturbances can break pushing contact and cause large errors, and prior methods typically assume a stationary fluid environment—unrealistic in biological applications where flow rates vary by orders of magnitude. The proposed approach combines the sample efficiency of MPC with the robustness of reinforcement learning: a learned residual policy corrects for time-varying flow disturbances while the underlying MPC preserves safety guarantees. The authors, including Zhuoqun Zhang, Hsiu-Chin Lin, and Aaron T. Becker, demonstrate the method on a microrobotic cell pushing task, achieving robust performance across a wide range of flow conditions. This work is relevant to researchers in microrobotics, reinforcement learning, and biomedical manipulation.

Paper Overview

Field: Machine Learning
Authors: Zhuoqun Zhang, Hsiu-Chin Lin, Megan Justice, Lyle Muller, Muhammad Suhail Saleem, Owen Schwartz, Max B. Wang, Aaron T. Becker
arXiv: 2603.05448

Abstract (translated)

Contact-rich micromanipulation in microfluidic flow is challenging because small disturbances can break pushing contact and induce large pushing errors. Previous approaches often assume stationary fluid environments, which limits their effectiveness in biological applications like cell manipulation, where flow rates can vary by orders of magnitude. To address this, the authors propose a Residual Reinforcement Learning Model Predictive Control (RL-MPC) framework that combines the sample efficiency of MPC with the robustness of reinforcement learning. The approach learns a residual policy that corrects for time-varying flow disturbances while maintaining the safety guarantees of MPC. The method is demonstrated on a microrobotic cell pushing task, achieving robust performance across a wide range of flow conditions.

Key Takeaways

Problem: Microflow disturbances destabilize contact pushing in micromanipulation; existing controllers assume static fluid environments.
Approach: Residual RL layered on top of MPC — RL learns to compensate for unmodeled, time-varying flow, while MPC retains safety constraints.
Result: Robust microrobotic cell pushing under wide-ranging flow conditions.

---

*Auto-collected on 2026-03-07.*