Learning Constraint-Aware Dynamical Systems from Human Demonstrations for Constrained Manipulation Tasks
Soyoun Sung, Keehoon Kim
AI summary
Problem
Existing learning-from-demonstration methods struggle to incorporate critical task-specific constraints like grasp locations and motion restrictions, leading to unreliable execution in highly constrained, tool-held scenarios.
Approach
The authors propose a constraint-aware dynamical system framework that uses Gaussian Process Regression to automatically identify graspable configurations, critical motion routes, and via-points directly from demonstration data, then embeds these as activation conditions for velocity generation.
Key results
- Automatically extracts graspable configurations and critical motion routes from demonstration data
- Models task-specific constraints (grasp-point, via-point, and constrained trajectories) using Gaussian Process Regression
- Enables reversible motion constraints through Lie algebra representation
- Achieves significantly higher task success rates than state-of-the-art methods on a 7-DoF manipulator and real-world daily-life tasks
Why it matters
Provides a reliable, adaptable framework for executing highly constrained robotic tasks in unstructured environments, bridging the gap between human demonstration and practical tool-use applications.
Abstract
Learning from demonstration (LfD) enables robots to acquire new skills from human examples without ex- plicit programming. Dynamical system (DS)-based approaches, in particular, have shown robustness to disturbances and adaptability in unstructured environments. However, existing methods often fail to incorporate task-specific constraints—such as grasp locations, execution starting points, or motion restric- tions—that are critical for reliable execution. This limitation becomes especially problematic in tool-use scenarios, where both the environment and the grasped tool impose strict restric- tions on feasible motions. To address this challenge, we propose a novel constraint-aware DS framework that automatically extracts and encodes task-specific constraints directly from demonstration data. The key idea is that task-critical configu- rations, repeatedly observed across successful demonstrations, can be identified and modeled as essential regions for task success using Gaussian Process Regression. By embedding these constraints, the proposed method generates motions that remain robust to environmental variations and tool-induced limitations. Experiments with a 7-DoF robotic manipulator demonstrate that our framework significantly improves task success rates over state-of-the-art methods. Real-world evaluations on daily- life tasks, such as dishware collection, further confirm its practicality and potential for real-world robotic applications.