Active Perception for Deformable Linear Objects Stiffness Estimation
Chengxiao Dong, Alessio Caporali, Hongyu Lan, Gianluca Palli
AI summary
Problem
Estimating the bending stiffness of Deformable Linear Objects (DLOs) in unconstrained environments is hindered by a lack of informative interaction data, as traditional methods rely on impractical fixed fixtures or passive probing that fails to sufficiently excite the object.
Approach
The authors reframe stiffness estimation as an active perception problem, training a Proximal Policy Optimization agent to autonomously learn adaptive manipulation strategies that dynamically excite the cable and maximize parameter identifiability.
Key results
- Fixed-end boundary constraints yield the highest estimation accuracy (MAE 0.0024)
- RL agent achieves MAE of 0.0192 and RMSE of 0.0252 on unconstrained cables
- Boundary conditions and grasp locations drastically govern stiffness identifiability
- Active exploration prevents estimator overfitting in stochastic manipulation sequences
Why it matters
Provides a practical, fixture-free paradigm for robust material property estimation in real-world robotic manipulation tasks.
Abstract
Estimating the stiffness of Deformable Linear Objects (DLOs) is crucial for robust manipulation. Inferring this hidden property depends heavily on the physical interaction strategy. Through a 1D CNN-based analysis of predefined probing modes, we first demonstrate that boundary constraints and grasp locations drastically alter stiffness identifiability. While fixed-end setups yield highly informative responses, they are rarely practical in unconstrained tasks. Consequently, we move beyond manual heuristics and reframe DLO parameter identification as an active perception problem. We propose a Reinforcement Learning (RL) framework that autonomously learns informative interaction strategies for free cables. By coupling a Proximal Policy Optimization (PPO) agent with a trajectory-aware estimator, the system dynamically excites the DLO to extract stiffness from diverse, stochastic manipu- lation sequences. Achieving a Mean Absolute Error (MAE) of 0.0192, our approach provides a robust, active paradigm that overcomes the limitations of static probing in unconstrained environments.