[Paper Review] Reactive Task and Motion Planning under Temporal Logic Specifications

Shen Li, Daehyung Park+3 more|arXiv (Cornell University)|Mar 26, 2021

AI-based Problem Solving and Planning27 references9 citations

TL;DR

This paper presents a reactive task-and-motion planning (TAMP) framework that enables robots to efficiently replan under human interventions using linear temporal logic (LTL) specifications and behavior trees. By leveraging experience-based incremental graph search and dynamically reconfigurable behavior trees, the system minimizes replanning overhead and ensures seamless execution, achieving superior time efficiency and motion cost-effectiveness in both simulation and real-world UR5 robot experiments.

ABSTRACT

We present a task-and-motion planning (TAMP) algorithm robust against a human operator's cooperative or adversarial interventions. Interventions often invalidate the current plan and require replanning on the fly. Replanning can be computationally expensive and often interrupts seamless task execution. We introduce a dynamically reconfigurable planning methodology with behavior tree-based control strategies toward reactive TAMP, which takes the advantage of previous plans and incremental graph search during temporal logic-based reactive synthesis. Our algorithm also shows efficient recovery functionalities that minimize the number of replanning steps. Finally, our algorithm produces a robust, efficient, and complete TAMP solution. Our experimental results show the algorithm results in superior manipulation performance in both simulated and real-world tasks.

Motivation & Objective

• Address the challenge of maintaining task completion under human interventions that invalidate precomputed robot plans.
• Reduce computational overhead during replanning by reusing prior planning experience and incrementally expanding search spaces.
• Enable seamless execution during environmental changes by integrating dynamically reconfigurable behavior trees with LTL-based planning.
• Achieve robust, efficient, and complete task-and-motion planning in human-robot collaborative scenarios with minimal replanning.
• Demonstrate effectiveness in both simulated environments and real-world robotic manipulation tasks using a UR5 platform.

Proposed method

• Formalize task objectives using linear temporal logic (LTL) to ensure correctness-by-design and support complex temporal specifications.
• Construct a partially expanded product automaton (PA) for efficient state-space exploration, avoiding full graph reconstruction after each intervention.
• Implement experience-based A* search that reuses previously computed paths and motion costs to accelerate replanning under environmental changes.
• Integrate a behavior tree (BT) with conditional subtree activation to enable runtime policy updates without halting execution.
• Dynamically reconfigure BT nodes based on environmental observations to maximize recovery range and reduce reliance on full replanning.
• Combine symbolic task planning with geometric motion-cost optimization to generate semi-optimal, executable action sequences.

Experimental results

Research questions

• RQ1 How can a robot efficiently replan its task and motion sequence when human interventions alter the environment, such as relocating or adding objects?
• RQ2 To what extent can prior planning experience reduce the computational cost of replanning under dynamic environmental changes?
• RQ3 Can behavior trees with conditional subtree structures effectively minimize the need for full replanning during task execution?
• RQ4 How does the proposed method compare to baseline approaches in terms of replanning time, motion cost, and completion efficiency?
• RQ5 Can the framework ensure task completion and seamless execution in real-world human-robot interaction scenarios?

Key findings

• The experience-based A* search reduced total replanning time to 0.69±0.43 seconds for object relocation, compared to 33.50±15.26 seconds with standard A* and 30.40±13.06 seconds with Dijkstra.
• For object removal, the experience-based method achieved a total replanning time of 0.10±0.02 seconds, significantly lower than 52.57±8.18 seconds (A*) and 60.53±9.16 seconds (Dijkstra).
• In object addition scenarios, the experience-based approach reduced total replanning time to 20.61±0.86 seconds, compared to 89.60±8.73 seconds (A*) and 116.16±16.35 seconds (Dijkstra).
• The system achieved 100% success rate across all 30 trials for each environmental change type, demonstrating robustness and completeness.
• Completion time was significantly reduced with the experience-based method: 466.50±34.80 seconds for relocation, 394.48±13.14 seconds for removal, and 647.00±24.67 seconds for addition, outperforming all baselines.
• The partial graph construction method scaled efficiently beyond 1,250 states, while full graph construction degraded exponentially, confirming its superiority in large-scale scenarios.