This paper from researchers at ETH Zurich presents a significant step forward in soft robotics: a spherical soft arm capable of catching a ball thrown by a human. While rigid robots have mastered ball-catching for decades, the inherent "squishiness" and complex material dynamics of soft robots usually make such high-precision, high-speed tasks nearly impossible.
TL;DR
The authors implemented an offset-free Model Predictive Control (MPC) scheme that compensates for the unpredictable material relaxation and modeling errors inherent in soft robots. By augmenting a linear model with real-time disturbance estimation, they reduced tracking errors by 35% and achieved a 91% success rate in autonomous ball catching.
What Problem Does This Paper Address?
Soft robots are prized for their safety and compliance, but they are notoriously difficult to control with precision. Unlike rigid robots, soft materials exhibit "material relaxation" and non-linear behaviors that are hard to capture in the compact mathematical models needed for real-time control.
Standard Model Predictive Control (MPC) predicts future behavior based on a model, but if that model is even slightly off—or if there are external disturbances—the robot will suffer from a "steady-state offset." In plain English: the robot will consistently miss its target by a few degrees. For a task like catching a ball, where the window of success is measured in centimeters and milliseconds, these offsets are fatal.
Methods
The researchers combined several clever engineering approaches to bridge the gap between soft-body unpredictability and high-precision control:
- Control Allocation & Decoupling: The arm is moved by three pneumatic bellows. 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 them to treat the complex arm as two independent, simpler spring-damper systems.
- Offset-free MPC: This is the core innovation. They took a standard linear state-space model and augmented it with a "disturbance state." A steady-state Kalman Filter estimates this disturbance in real-time. The MPC then solves an optimization problem that doesn't just aim for the target, but aims for a target adjusted to cancel out the estimated disturbance.
- Ball Trajectory Prediction: An Extended Kalman Filter (EKF) was used to track the ball via a motion capture system, accounting for gravity and aerodynamic drag to predict the exact intersection point with the robot's workspace.
- Great Circle Motion Planning: To move the arm to the catch point, the controller plans a path along the "great circle" (the shortest path on a sphere). This ensures smooth, high-speed movement without the jerky transitions that might excite unwanted oscillations in the soft material.
Datasets
The authors did not use a pre-existing public dataset; instead, they generated custom data for System Identification:
- Training Data: They applied sinusoidal signals (0.5 Hz to 5 Hz) and step inputs to the arm to identify the parameters for their linear model.
- Evaluation Trajectories: The controller was tested against random ramp trajectories, soft step changes, and sinusoidal paths.
- Ball Catching: The final evaluation involved 68 physical trials of a human throwing a ball toward the robot.
Key Insights and Ideas
- Linearity is Sufficient: One of the most surprising takeaways is that you don't necessarily need a complex, non-linear "soft physics" model or a deep neural network to control a soft robot. A simple linear model, when paired with a robust disturbance estimator, can outperform more complex models that lack offset compensation.
- Disturbance as a "Catch-all": By treating material relaxation and modeling errors as a generic "disturbance," the authors bypassed the need to mathematically define every complex property of the silicone and fabric used in the arm.
- The Importance of Path Geometry: Planning the trajectory on the surface of the sphere (the great circle) was critical. It allowed the robot to reach an angular velocity of 240 deg/s—essential for catching—while maintaining the stability of the soft structure.
Key Results
- Error Reduction: The offset-free MPC achieved an average Root Mean Square Error (RMSE) of 1.8°, a 35% improvement over standard MPC (2.8°).
- Catching Success: In 68 trials, the robot successfully caught the ball 62 times (91% success rate).
- High Precision: For successful catches, the ball landed an average of only 9.7 mm from the center of the net (which has a 31 mm radius).
- Robustness: The system remained effective even when ventilation fans were used to create wind gusts, as the estimator constantly re-adjusted the predicted intersection point.
Limitations and Future Work
- Small Angle Assumption: The decoupling of the $\alpha$ and $\beta$ dynamics is an approximation that holds best at small angles. Accuracy may degrade at the extreme edges of the robot's range of motion.
- Static Stiffness: In this study, the joint stiffness was kept at a constant value. The authors note that future work will involve using the offset-free MPC to actively control the robot's stiffness, which could be useful for tasks like interacting with humans or handling heavier objects.
- Hardware Specifics: The current success relies on a high-speed (200 Hz) infrared motion capture system; transitioning this to on-board vision would be a significant next step for real-world deployment.
Summary updated