Offset-free Model Predictive Control: A Ball Catching Application with a Spherical Soft Robotic Arm

Yaohui Huang, Matthias Hofer, Raffaello D'Andrea

Introduction

Researchers at ETH Zurich have demonstrated a significant advancement in soft robotics: a spherical soft arm capable of catching a ball thrown by a human. While rigid robots have mastered high-speed manipulation for decades, the inherent compliance and complex material dynamics of soft robots typically make high-precision tasks like ball catching extremely difficult.

The Challenge of Soft Robot Control

Soft robots are valued for their safety and adaptability, but they are notoriously difficult to control with precision. Unlike their rigid counterparts, soft materials exhibit "material relaxation" and non-linear behaviors that are hard to capture in the compact mathematical models required for real-time control.

Standard Model Predictive Control (MPC) relies on a predictive model. If this model is inaccurate—due to unmodeled dynamics or external disturbances—the robot suffers from a "steady-state offset," consistently missing its target. For ball catching, where success is measured in centimeters and milliseconds, these offsets are unacceptable.

Methodology

The researchers implemented a multi-layered control strategy to bridge the gap between soft-body unpredictability and high-precision requirements.

Control Allocation and Decoupling

The arm is actuated by three pneumatic bellows. To simplify control, the authors used a bijective mapping to transform these three pressure inputs into two decoupled orientation angles ($\alpha$ and $\beta$) and one stiffness parameter. This allowed the complex arm to be treated as two independent spring-damper systems.

Schematic of the soft arm showing the coordinate system ($\alpha, \beta$) and the layout of the three pneumatic actuators (A, B, C).

Offset-free MPC

The core innovation is an offset-free MPC scheme. A standard linear state-space model is augmented with a "disturbance state," estimated in real-time by a steady-state Kalman Filter. The MPC then solves an optimization problem that aims for a target adjusted to cancel out this estimated disturbance, effectively compensating for material relaxation and modeling errors.

Trajectory Planning and Prediction

An Extended Kalman Filter (EKF) tracks the ball via a motion capture system, accounting for gravity and drag to predict the intersection point. To reach this point, the controller plans a path along the "great circle"—the shortest path on a sphere—ensuring smooth, high-speed movement (up to 240 deg/s) without exciting unwanted oscillations.

Key Results

The implementation of offset-free MPC led to substantial performance gains over standard MPC.

  • Error Reduction: The offset-free scheme achieved an average Root Mean Square Error (RMSE) of 1.8°, representing a 35% improvement over standard MPC (2.8°).
  • Catching Success: In 68 physical trials involving human throws, the robot successfully caught the ball 62 times (91% success rate).
  • High Precision: Successful catches landed an average of only 9.7 mm from the center of the net.
Comparison of tracking performance between standard MPC and Offset-free MPC (OSFMPC). The OSFMPC (orange) significantly reduces the steady-state offset compared to standard MPC (blue) relative to the setpoint (dashed black).

Conclusion and Future Work

This research demonstrates that complex "soft physics" models or deep neural networks are not always necessary for high-performance soft robot control. A simple linear model, when paired with robust disturbance estimation, can achieve remarkable precision. Future work will explore active stiffness control and transitioning from external motion capture to on-board vision systems for real-world deployment.

Summary updated

BibTeX
@inproceedings{huang2021offset,
  author = {Yaohui Huang and Matthias Hofer and Raffaello D'Andrea},
  title = {{Offset-free Model Predictive Control: A Ball Catching Application with a Spherical Soft Robotic Arm}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2021},
  doi = {10.1109/iros51168.2021.9636608},
}
Figure