Real-Time Trajectory Adaptation for Tossing Robots Using Soft Switching Multiple Model Predictive Control

Babak Mehdizadeh Gavgani, Thijs Van Hauwermeiren, Arash Farnam, Majid Ghorbani, Jeroen D. M. De Kooning, Guillaume CrevecoeurJ. Intell. Robotic Syst 2025 · DOI

Introduction

Robotic tossing has emerged as a high-speed alternative to traditional pick-and-place operations, offering the potential to significantly increase throughput in logistics and automated sorting. Unlike standard manipulation, tossing allows a robot to release an object toward a target, effectively extending its workspace and reducing cycle times. However, a major hurdle remains: most existing tossing systems rely on predefined, static trajectories. If a target container moves or a new destination is assigned after the motion has begun, these systems typically fail.

To achieve true flexibility, a robot must be able to adapt its trajectory in real-time. This requires a control strategy capable of handling complex nonlinear dynamics and strict hardware constraints—such as joint velocity and torque limits—within milliseconds.

The Challenge of Dynamic Tossing

The standard approach for handling such constraints in dynamic environments is Nonlinear Model Predictive Control (NLMPC). While NLMPC is highly accurate, it is computationally expensive and often unpredictable in its execution time. For a high-speed tossing motion that may last only a second, the tens of seconds required for NLMPC calculations are impractical. Conversely, Linear MPC (LMPC) is computationally efficient but lacks the range to manage the robot's full nonlinear workspace effectively.

Proposed Method: SSM-MPC

This paper introduces Soft Switching Multiple Model Predictive Control (SSM-MPC), a strategy designed to bridge the gap between the speed of linear control and the robustness of nonlinear methods. The SSM-MPC framework operates through three key components:

  1. Multilinear Model Bank: Instead of relying on a single, complex nonlinear model, the authors construct a "bank" of simplified linear models. Each model is optimized for a specific segment of the robot's operational space.
  2. Gap Metric Optimization: To ensure computational efficiency, the system must not be overloaded with models. The authors utilize the "gap metric"—a mathematical measure of dynamic similarity—to prune an initial set of 300 models down to just five essential operating points.
  3. Soft Switching Mechanism: When a target change is detected mid-toss, the system identifies the current state and activates the most appropriate linear MPC from the bank. This allows for a "soft" transition between trajectories, ensuring the robot reaches the new target without violating physical constraints.
Conceptual diagram of trajectory switching between target locations A and B.

As shown in the figure above, the switching trajectory ($\tau_{sw}$) smoothly guides the robot from an initial path (e.g., toward $B_1$) to a new path (toward $B_2$) starting from the switching point ($SW$).

Experimental Validation

The proposed SSM-MPC was validated using a 6-degree-of-freedom (6-DOF) UR5 industrial robot arm equipped with a pneumatic suction gripper. The experimental task involved tossing a 40mm ping-pong ball into three distinct containers (A, B, and C) positioned at varying angles (0°, -12°, and -20°) relative to the robot's base.

3D trajectories generated for three different target containers (A, B, and C).

The controller was implemented in C++ using the Acados toolkit, with an NLMPC solver (IPOPT) serving as the performance benchmark. The study focused on the system's ability to switch targets mid-motion while adhering to the UR5's joint velocity and torque limits.

Results and Discussion

The results demonstrate a dramatic improvement in computational efficiency. The SSM-MPC achieved a 1300x reduction in computational load compared to the NLMPC benchmark. Specifically, the average computation time for the proposed method was a mere 40ms (0.04 seconds), whereas the NLMPC required an average of 52 seconds—a duration far exceeding the 1.2-second window of the actual tossing motion.

Beyond speed, the system maintained high precision. Experimental data showed that the robot could switch targets with minimal tracking error, keeping joint angle deviations below $2 \times 10^{-4}$ radians. Furthermore, the "soft switching" approach ensured that all transitions remained within the safe operating limits of the hardware, avoiding the jerky motions often associated with switched control systems.

Conclusion

The SSM-MPC strategy provides a viable path for deploying high-speed, adaptive robotic tossing in dynamic industrial environments. By leveraging a pruned bank of linear models and the gap metric for optimization, the system achieves real-time performance without sacrificing the ability to handle complex constraints.

Future work may involve refining the model bank to handle a wider variety of toss types and integrating aerodynamic flight models to improve accuracy for objects more susceptible to air drag. Additionally, exploring optimization-based selection for transition points could further enhance the energy efficiency of the switching process.

Summary updated

BibTeX
@article{gavgani2025real,
  author = {Babak Mehdizadeh Gavgani and Thijs Van Hauwermeiren and Arash Farnam and Majid Ghorbani and Jeroen D. M. De Kooning and Guillaume Crevecoeur},
  title = {{Real-Time Trajectory Adaptation for Tossing Robots Using Soft Switching Multiple Model Predictive Control}},
  journal = {J. Intell. Robotic Syst},
  year = {2025},
  doi = {10.1007/s10846-025-02297-2},
}
Figure