Trajectory Optimization for Manipulation of Deformable Objects: Assembly of Belt Drive Units

Shiyu Jin, Diego Romeres, Arvind Ragunathan, Devesh K. Jha, Masayoshi Tomizuka

Abstract

The assembly of belt drive units presents a significant challenge in robotics due to the complex, infinite-dimensional dynamics of deformable objects. This paper proposes a trajectory optimization framework that simplifies the belt into a 3D keypoint representation and utilizes Mathematical Programming with Complementarity Constraints (MPCC) to automate the assembly process. By treating contact as a continuous mathematical constraint, the system avoids the need for manual mode scheduling. Experimental results in both high-fidelity simulation and real-world robotic setups demonstrate an 85% success rate and computational efficiency suitable for long-horizon planning.

Introduction

Manipulating deformable objects like belts, ropes, or cables is considered a "grand challenge" in robotics. Unlike rigid bodies, a belt has infinite degrees of freedom, and its dynamics shift drastically between slack and taut states. Traditionally, planning these tasks requires a "hybrid" approach where engineers manually define different contact modes (e.g., slack vs. touching a pulley) and plan transitions between them. This process is labor-intensive and often fragile.

This research aims to automate the trajectory planning for wrapping a polyurethane belt around two pulleys—a task highlighted in the 2018 World Robot Summit. The goal is to create a generalizable framework that handles shifting physics without pre-defined contact sequences.

The experimental setup featuring a 6-DOF robot arm, F/T sensor, and the belt-pulley assembly task.

Proposed Method

The authors introduce a two-pronged approach to manage the complexity of deformable manipulation:

1. 3D Keypoint Representation

Instead of modeling the entire continuous geometry of the belt, the system represents it using two primary keypoints: $K_1$ (the point held by the robot gripper) and $K_2$ (the point furthest from the gripper). The connection between these points is modeled as a "virtual elastic cable," which drastically reduces the computational burden while capturing the essential physics of tension and slack.

The 3D keypoint representation (K1, K2) used to simplify the belt's complex geometry into a manageable model.

2. MPCC Formulation

To handle the "on/off" nature of contact and tension, the researchers utilize Mathematical Programming with Complementarity Constraints (MPCC). These constraints act as mathematical switches (e.g., either the contact force is zero, or the distance to the pulley is zero). This allows the optimization solver, Ipopt, to autonomously discover the necessary contact sequence—determining when to touch the pulley and when to pull the belt taut—without human intervention.

The task is decomposed into two subtasks: $S_1$ (wrapping the belt around the first pulley until taut) and $S_2$ (rotating around the first pulley to hook the belt onto the second).

Experimental Evaluation

The framework was evaluated using both simulation and physical hardware:

  • Simulation: The MuJoCo physics engine was used to provide high-fidelity ground truth, modeling the belt as 41 linked capsules. The system was tested across four distinct scenarios with varying pulley positions.
  • Real-World Setup: A physical testbed was constructed using a 6-DOF Fanuc LR-Mate 200iD robot arm equipped with an ATI force/torque (F/T) sensor and 3D-printed pulleys.

Key Results and Insights

  • High Success Rate: In 40 simulation trials, the system achieved an 85% success rate (34/40). Failures were primarily due to sampled goal positions being slightly out of reach for a successful wrap.
  • Computational Efficiency: Trajectories with 600 time steps were computed in approximately 36 seconds, demonstrating feasibility for complex long-horizon planning.
  • Physical Validation: Real-world experiments confirmed that the physical belt tension closely matched the "virtual" tension predicted by the MPCC model, validating the abstraction of the belt as a virtual elastic cable.
  • Elasticity as a Tool: The optimizer specifically leveraged the belt's elasticity in subtask $S_2$, encouraging tension to ensure the belt remained secure on the first pulley while reaching for the second.

Conclusion and Future Work

While the current framework successfully automates belt assembly, it operates in an open-loop fashion, meaning it cannot adjust to mid-motion slips or pulley shifts. Future work will focus on integrating a feedback controller to handle real-time perturbations. Additionally, the researchers aim to automate keypoint selection for more complex deformable shapes and extend the method to more intricate routing tasks, such as engine serpentine belts.

Summary updated

Abstract

This paper presents a novel trajectory optimization formulation to solve the robotic assembly of the belt drive unit. Robotic manipulations involving contacts and deformable objects are challenging in both dynamic modeling and trajectory planning. For modeling, variations in the belt tension and contact forces between the belt and the pulley could dramatically change the system dynamics. For trajectory planning, it is computationally expensive to plan trajectories for such hybrid dynamical systems as it usually requires planning for discrete modes separately. In this work, we formulate the belt drive unit assembly task as a trajectory optimization problem with complementarity constraints to avoid explicitly imposing contact mode sequences. The problem is solved as a mathematical program with complementarity constraints (MPCC) to obtain feasible and efficient assembly trajectories. We validate the proposed method both in simulations with a physics engine and in real-world experiments with a robotic manipulator.
BibTeX
@inproceedings{jin2021trajectory,
  author = {Shiyu Jin and Diego Romeres and Arvind Ragunathan and Devesh K. Jha and Masayoshi Tomizuka},
  title = {{Trajectory Optimization for Manipulation of Deformable Objects: Assembly of Belt Drive Units}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2021},
  url = {https://openreview.net/forum?id=SDAnboAF5k_},
}
Figure