A Behavior Tree and Dynamic Motion Primitive-Based Framework for Learning and Executing Robotic Tasks From Demonstration
IEEE Transactions on Automation Science and Engineering, vol. 23, pp. 5022–5035
Abstract
Learning from Demonstrations(LfD) enables robots to acquire complex skills by observing human behavior, significantly reducing the need for explicit programming. However, applying LfD in industrial settings remains challenging due to limited demonstrations, variability in task executions, and the need to generalize across diverse scenarios. To address these issues, this paper presents a learning based hierarchical task and motion planning framework that integrates Behavior Trees (BT) for high-level task sequencing and Dynamic Motion Primitives (DMP) for low-level motion generation. Demonstration trajectories are segmented and actions are generated using an agentcentric, state-augmented segmentation strategy. Subsequently, relevant features are automatically extracted to define the pre-and post-conditions for each action for the construction of a modular BT. For motion execution, DMP are enhanced with a recovery mechanism for adaptive, obstacle-aware reproduction. A backchaining mechanism is also introduced for BT extension. Validation was performed through simulation and real-world experiments on multiple tasks. Comparative results demonstrate that the proposed method outperforms existing LfD and planning baselines in task success rate, efficiency, and motion smoothness, highlighting its potential for flexible and scalable automation.
Authors 3
-
Affiliation as printed
Institute of Mechanism Theory, Machine Dynamics and Robotics, RWTH Aachen University, Aachen, Germany
-
Affiliation as printed
Institute of Mechanism Theory, Machine Dynamics and Robotics, RWTH Aachen University, Aachen, Germany
-
Affiliation as printed
Department of Mechanical Engineering, The University of Tokyo, Tokyo, Japan
Cited by 0 stored of 0
No patents citing this paper on Lens.org (checked 2026-10-06).