Learning Problem Decomposition for Efficient Sequential Multi-Object Manipulation Planning (resubmitted)
Yan Zhang, Teng Xue, Amirreza Razmjoo, Sylvain Calinon
AI summary
Problem
Conventional Task and Motion Planning (TAMP) solvers face exponential planning time growth as task horizons and object counts increase, hindering real-time closed-loop replanning under disturbances.
Approach
The framework extracts mandatory subgoal sequences from demonstrations using sequential pattern mining, trains a graph neural network to predict object importance as a computational distance metric, and prunes irrelevant objects during replanning to accelerate classical TAMP solvers.
Key results
- Unsupervised extraction of mandatory subgoal sequences from demonstrations via sequential pattern mining
- Graph neural network trained to predict object importance as a computational distance metric for subgoal selection
- Object reduction mechanism that prunes irrelevant objects during replanning while preserving solution completeness
- Validated efficiency gains across three sequential multi-object manipulation benchmarks under dynamic disturbances
Why it matters
Enables real-time, robust robotic manipulation in complex, dynamic environments where traditional TAMP solvers are computationally prohibitive.
Abstract
We present an efficient task and motion replan- ning approach for sequential multi-object manipulation in dynamic environments. Conventional Task And Motion Plan- ning (TAMP) solvers experience an exponential increase in planning time as the planning horizon and number of objects grow, limiting their applicability in real-world scenarios. To address this, we propose learning problem decompositions from demonstrations to accelerate TAMP solvers. Our approach consists of three key components: goal decomposition learning, computational distance learning, and object reduction. Goal decomposition identifies the necessary sequences of states that the system must pass through before reaching the final goal, treating them as subgoal sequences. Computational distance learning predicts the computational complexity between two states, enabling the system to identify the temporally clos- est subgoal from a disturbed state. Object reduction mini- mizes the set of active objects considered during replanning, further improving efficiency. We evaluate our approach on three benchmarks, demonstrating its effectiveness in improving replanning efficiency for sequential multi-object manipula- tion tasks in dynamic environments. Accompanying website: https://sites.google.com/view/problem-decomposition-ral