HandMeThat: Human-Robot Communication in Physical and Social Environments

TL;DR

HandMeThat is a comprehensive benchmark for evaluating human-robot communication in physical and social contexts, emphasizing understanding ambiguous instructions through physical and social cues, with initial models showing limited performance.

Contribution

The paper introduces HandMeThat, a new benchmark dataset for holistic evaluation of instruction understanding in physical and social environments, including a textual interface and baseline evaluations.

Findings

Baseline models perform poorly, indicating room for improvement.

The benchmark covers physical and social cues in human-robot interactions.

HandMeThat contains 10,000 episodes of human-robot interaction data.

Abstract

We introduce HandMeThat, a benchmark for a holistic evaluation of instruction understanding and following in physical and social environments. While previous datasets primarily focused on language grounding and planning, HandMeThat considers the resolution of human instructions with ambiguities based on the physical (object states and relations) and social (human actions and goals) information. HandMeThat contains 10,000 episodes of human-robot interactions. In each episode, the robot first observes a trajectory of human actions towards her internal goal. Next, the robot receives a human instruction and should take actions to accomplish the subgoal set through the instruction. In this paper, we present a textual interface for our benchmark, where the robot interacts with a virtual environment through textual commands. We evaluate several baseline models on HandMeThat, and show that both offline and online reinforcement learning algorithms perform poorly on HandMeThat, suggesting significant room for future work on physical and social human-robot communications and interactions.