Interpretable Responsibility Sharing as a Heuristic for Task and Motion Planning

This article introduces a novel heuristic for Task and Motion Planning (TAMP) named Interpretable Responsibility Sharing (IRS), which enhances planning efficiency in domestic robots by leveraging human-constructed environments and inherent biases. Utilizing auxiliary objects (e.g., trays and pitchers), which are commonly found in household settings, IRS systematically incorporates these elements to simplify and optimize task execution. The heuristic is rooted in the novel concept of Responsibility Sharing (RS), where auxiliary objects share the task’s responsibility with the embodied agent, dividing complex tasks into manageable sub-problems. This division not only reflects human usage patterns but also aids robots in navigating and manipulating within human spaces more effectively. By integrating Optimized Rule Synthesis (ORS) for decision-making, IRS ensures that the use of auxiliary objects is both strategic and context-aware, thereby improving the interpretability and effectiveness of robotic planning. Experiments conducted across various household tasks demonstrate that IRS significantly outperforms traditional methods by reducing the effort required in task execution and enhancing the overall decision-making process. This approach not only aligns with human intuitive methods but also offers a scalable solution adaptable to diverse domestic environments. Code is available at https://github.com/asyncs/IRS.

💡 Research Summary

**
The paper introduces Interpretable Responsibility Sharing (IRS), a novel heuristic designed to improve task and motion planning (TAMP) for domestic robots by explicitly exploiting the “human‑centric environmental bias” that naturally arises in homes. The authors observe that everyday objects such as trays, pitchers, or even stationary robots are deliberately placed by humans to facilitate recurring tasks. Rather than learning these biases indirectly through demonstrations or end‑to‑end deep networks, IRS treats these auxiliary objects as active participants that can share part of the robot’s workload—a concept they call Responsibility Sharing (RS).

IRS operates in three tightly coupled stages. First, Counterfactual Plan Generation (CPG) creates a labeled dataset for each task instance by synthesizing two plans: a baseline plan that manipulates the target directly (W = 0) and a counterfactual plan that employs an auxiliary object (W = 1). Physical effort metrics (e.g., total displacement, energy consumption) are computed for both plans, and the Individual Treatment Effect (ITE) is defined as the cost difference. A negative ITE indicates that using the auxiliary object reduces effort, and this outcome becomes the supervision signal for the next stage.

Second, Optimized Rule Synthesis (ORS) learns interpretable first‑order‑logic rules from the CPG data. ORS combines a Rule‑Based Representation Learner (RRL) that maps high‑dimensional continuous features (object poses, distances, robot joint states) into discrete predicates, with a Correlation‑and‑Order‑Aware Rule Learning (CARL) module that selects a compact, high‑information rule set. The selection objective, called the Balance Score, simultaneously maximizes predictive accuracy and ensures that each rule is grounded in a measurable physical benefit (e.g., “if a tray is within 0.5 m of the source, then group items on the tray”). The resulting rule base is human‑readable, concise, and directly reflects the inductive bias present in the environment.

Third, the IRS heuristic is applied at inference time. When the planner receives a new problem, it queries the learned rule base. If a rule fires, the planner decomposes the original task into sub‑goals that correspond to the shared responsibility (e.g., “place objects on tray”, “move tray to destination”, “unload objects”). Each sub‑goal is then solved independently using any underlying TAMP engine—here the authors employ Logic‑Geometric Programming (LGP) for continuous trajectory optimization and Multi‑Bound Tree Search (MBTS) for symbolic branching. If no rule applies, the planner falls back to the standard search strategy. Consequently, IRS functions as a meta‑heuristic that is orthogonal to, and compatible with, existing TAMP solvers.

The authors evaluate IRS on three representative domestic scenarios: (1) serving, where multiple items must be delivered to different locations; (2) pouring, where a liquid source must be distributed to several containers; and (3) hand‑over, where a stationary robot acts as an auxiliary object for transferring items between mobile agents. Across 120 benchmark instances, IRS reduces planning time by 30‑45 % and total motion effort by 20‑35 % compared with baseline LGP/MBTS without IRS. Notably, in the hand‑over task the method demonstrates that responsibility sharing can extend to multi‑robot contexts, with the stationary robot effectively becoming a “tool” that carries part of the workload.

A complementary human study asked participants to solve the same serving tasks without a robot. The majority reported naturally grouping items on a tray before carrying them, confirming that the rules learned by IRS align with human intuition. This alignment bolsters the claim that IRS not only improves computational efficiency but also enhances trust and transparency, which are crucial for domestic deployment.

Key contributions of the work are: (i) a systematic pipeline that extracts interpretable logical rules from counterfactual physical‑cost analysis, thereby making latent environmental biases explicit; (ii) the formulation of Responsibility Sharing as a planning meta‑heuristic that decomposes tasks into sub‑problems without altering the underlying TAMP algorithm; (iii) a novel rule synthesis framework (ORS) that balances rule compactness, predictive power, and physical grounding; and (iv) empirical evidence that the approach scales across different task families and even multi‑robot settings. The paper concludes by outlining future directions, including handling partially observable or dynamically changing environments, scaling the rule learning to richer object vocabularies, and integrating IRS with learning‑based motion planners for even greater adaptability.