[Paper Review] Levels of Integration between Low-Level Reasoning and Task Planning

Esra Erdem, Volkan Patoğlu+1 more|arXiv (Cornell University)|Jul 29, 2013

AI-based Problem Solving and Planning16 references3 citations

TL;DR

This paper proposes and evaluates four distinct strategies for integrating low-level geometric and temporal reasoning into high-level task planning in robotics, using answer set programming with external predicates. The key finding is that interleaved, on-demand integration (Int) offers the best performance and feasibility, while replanning (Repl) provides robustness when full integration is impractical.

ABSTRACT

We provide a systematic analysis of levels of integration between discrete high-level reasoning and continuous low-level reasoning to address hybrid planning problems in robotics. We identify four distinct strategies for such an integration: (i) low-level checks are done for all possible cases in advance and then this information is used during plan generation, (ii) low-level checks are done exactly when they are needed during the search for a plan, (iii) first all plans are computed and then infeasible ones are filtered, and (iv) by means of replanning, after finding a plan, low-level checks identify whether it is infeasible or not; if it is infeasible, a new plan is computed considering the results of previous low- level checks. We perform experiments on hybrid planning problems in robotic manipulation and legged locomotion domains considering these four methods of integration, as well as some of their combinations. We analyze the usefulness of levels of integration in these domains, both from the point of view of computational efficiency (in time and space) and from the point of view of plan quality relative to its feasibility. We discuss advantages and disadvantages of each strategy in the light of experimental results and provide some guidelines on choosing proper strategies for a given domain.

Motivation & Objective

• To systematically analyze levels of integration between discrete high-level reasoning and continuous low-level reasoning in hybrid robotic planning problems.
• To evaluate the trade-offs between computational efficiency (time and memory) and plan quality (feasibility) across different integration strategies.
• To provide practical guidelines for selecting appropriate integration strategies based on domain characteristics and computational constraints.
• To assess the impact of precomputation, on-demand checking, filtering, and replanning on performance in robotic manipulation and legged locomotion domains.

Proposed method

• Uses answer set programming (ASP) with the dlvhex solver to represent task planning problems and external predicates for low-level feasibility checks.
• Employs four integration strategies: (i) precomputation of all low-level checks before planning, (ii) on-demand low-level checks during search, (iii) plan generation followed by filtering of infeasible plans, and (iv) replanning after detecting infeasibility.
• Combines strategies in hybrid approaches, such as using precomputation for some checks and on-demand for others, to explore intermediate integration levels.
• Implements dedicated precomputation tools to reduce overhead and improve efficiency, especially in domains with large input spaces for low-level checks.
• Measures performance via runtime, memory usage, number of feasible and infeasible solutions generated, and solution enumeration efficiency.
• Uses robotic manipulation and legged locomotion as test domains to evaluate strategies under realistic constraints and complex feasibility checks.

Experimental results

Research questions

• RQ1 How do different levels of integration between high-level task planning and low-level geometric reasoning affect computational efficiency and plan feasibility?
• RQ2 What are the performance trade-offs between precomputation, on-demand checking, filtering, and replanning strategies in hybrid planning?
• RQ3 In which domains or conditions is precomputation more effective than on-demand integration, and vice versa?
• RQ4 How does the integration level impact the number of infeasible solutions generated before finding a feasible one?
• RQ5 Can hybrid strategies combining multiple integration levels improve robustness and efficiency compared to single-strategy approaches?

Key findings

• Interleaved, on-demand integration (Int) achieves the best runtime performance and generates the fewest infeasible solutions, with 621 feasible plans and only 38 infeasible candidates in the Robotic Manipulation domain.
• Precomputation (Pre) outperforms other methods in Robotic Manipulation when efficient precomputation tools are used, reducing total computation time from 1361 seconds (naive) to 238 seconds.
• Filtering (Filt) performs worst, generating 11,787 solution candidates versus only 622 feasible ones in Robotic Manipulation, indicating poor search guidance.
• Replanning (Repl) is more robust than Filt, successfully solving all 20 instances in Legged Locomotion, while Filt solved only 15 out of 20.
• In Legged Locomotion, Repl generates 250 infeasible candidates compared to 68 feasible solutions, suggesting higher infeasibility rates due to large input space for balance checks.
• Combining Pre with other strategies improves efficiency, especially in domains with high-dimensional inputs, but precomputation becomes infeasible if low-level checks are too expensive to compute in advance.