Robust Planning for Human-Robot Joint Tasks with Explicit Reasoning on Human Mental State

Anthony Favier, Shashank Shekhar, Rachid Alami

TL;DR

This paper introduces a planning framework that allows robots to collaborate more effectively with humans by explicitly modeling the human's "mental state"—specifically, what the human knows versus what they don't. By tracking belief divergences and using "situation assessment," the robot can decide exactly when and what to communicate to keep a joint task on track, achieving a 100% success rate in complex scenarios where traditional planners often fail.

What Problem Does This Paper Address?

In human-robot collaboration (HRC), robots often fail because they make one of two flawed assumptions: either they treat the human as a controllable machine or they assume the human is "omniscient" (knowing everything the robot knows).

In the real world, humans have limited visibility and different perspectives. If a robot turns on a stove in the kitchen while the human is in the living room, the human doesn't know the stove is on. Traditional planners might lead to an "inactivity deadlock," where the human waits for the robot to do something that is already finished, or a "non-applicable action," where the human tries to perform a task they aren't ready for. The challenge is enabling the robot to "put itself in the human's shoes" to predict these misunderstandings and fix them through minimal, effective communication.

Methods

The authors build upon the HATP/EHDA (Human-Aware Task Planner / Emulating Human Decisions and Actions) framework, which uses Hierarchical Task Networks (HTNs). Their novel contributions include:

  • Place-Based Observability: They categorize world attributes as either Observable (OBS)—things an agent can see if they are in the same room (e.g., a stove being on)—or Inferrable (INF)—things that aren't visible even if you are there (e.g., whether salt has already been added to a pot).
  • Situation Assessment: The robot constantly updates its estimate of the human's belief state. If the human moves into a room, the robot "knows" the human will automatically update their beliefs about all OBS attributes in that room.
  • Minimal Communication Planning: When the robot detects a "relevant" belief divergence (one that would cause the human to make a mistake or stall), it uses a Breadth-First Search (BFS) to find the smallest set of facts it needs to tell the human to align their beliefs enough to complete the task. This prevents the robot from being "too wordy" or annoying.

Datasets

The researchers evaluated their approach using two custom-designed simulation domains:

  1. Cooking Pasta Domain: A scenario involving a kitchen and a living room where agents must coordinate turning on a stove, adding salt, and pouring pasta.
  2. Preparing Box Domain: A task where agents fill a box with balls and paste a sticker. It involves tracking "inferrable" counts (number of balls) and "observable" states (sticker presence).

The authors generated 512 different initial states for these domains, including 448 cases with divergent initial beliefs, resulting in a total of 2,048 plans for evaluation.

Key Insights and Ideas

  • Theory of Mind (ToM) in Planning: The core "aha" moment is that a robot shouldn't just plan the physical steps of a task; it must plan the informational state of its partner.
  • Communication as a Tool, Not a Constant: Communication is treated as a specific action triggered by necessity. By modeling what a human learns just by "looking around" (situation assessment), the robot can avoid redundant talking.
  • Myopic but Effective Reasoning: The current system is "myopic," meaning it aligns beliefs as soon as a divergence becomes relevant to the next possible actions. While simple, this proved remarkably robust in preventing deadlocks.
  • Handling Uncontrollability: Because the robot cannot "command" the human, it instead "creates circumstances" (through actions or words) that elicit the desired behavior from the human.

Key Results

The researchers compared their new solver against the "Old Solver" (standard HATP/EHDA):

Metric Old Solver (Avg) Our Solver (Avg)
Success Rate (S) 21.8% 100%
Failed: Non-applicable action (NA) 80.2% 0%
Failed: Inactivity Deadlock (IDL) 19.9% 0%
Plans requiring communication N/A 61.9%

The old solver failed nearly 80% of the time because it assumed the human knew things they couldn't possibly see. The new solver achieved a perfect success rate by identifying when the human was "in the dark" and proactively communicating or waiting for the human to observe the state themselves.

Limitations and Future Work

  • Myopic Reasoning: The system currently doesn't look far into the future to see if a small belief divergence now might cause a massive problem much later; it only fixes what is immediately relevant.
  • Deterministic Humans: The model assumes humans are rational and move deterministically. It doesn't yet account for humans who might be distracted, irrational, or forgetful.
  • Communication Costs: Currently, all communication actions are treated as having equal "cost." Future work could model the "social cost" of interrupting a human versus the benefit of the information.
  • Experimental Validation: The results are currently based on simulations. The authors plan to move to real-world human-robot experiments to substantiate their claims.

Summary updated

Abstract

We consider the human-aware task planning problem where a human-robot team is given a shared task with a known objective to achieve. Recent approaches tackle it by modeling it as a team of independent, rational agents, where the robot plans for both agents' (shared) tasks. However, the robot knows that humans cannot be administered like artificial agents, so it emulates and predicts the human's decisions, actions, and reactions. Based on earlier approaches, we describe a novel approach to solve such problems, which models and uses execution-time observability conventions. Abstractly, this modeling is based on situation assessment, which helps our approach capture the evolution of individual agents' beliefs and anticipate belief divergences that arise in practice. It decides if and when belief alignment is needed and achieves it with communication. These changes improve the solver's performance: (a) communication is effectively used, and (b) robust for more realistic and challenging problems.
BibTeX
@misc{favier2022robust,
  author = {Anthony Favier and Shashank Shekhar and Rachid Alami},
  title = {{Robust Planning for Human-Robot Joint Tasks with Explicit Reasoning on Human Mental State}},
  year = {2022},
  eprint = {2210.08879},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2210.08879},
}
Figure