Controlling Intent Expressiveness in Robot Motion with Diffusion Models

TL;DR

This work addresses legibility in robot motion by treating it as a spectrum of goal-inference expressiveness rather than a single optimal trajectory. It introduces an Information Potential Field ($IPF$) to compute continuous legibility scores and a two-stage diffusion pipeline (Intent-Aware Path Diffuser and Path-Guided Diffusion Policy) that generates legible yet executable trajectories conditioned on a desired legibility level $\ell(\xi)\in [-1,1]$. A Quality-Diversity trajectory dataset supports training across the legibility spectrum, with trajectories parameterized as cubic Bézier curves and scored by a potential-based measure. Experiments in 2D and 3D reaching tasks demonstrate controllable legibility across the spectrum while maintaining competitive performance to state-of-the-art baselines in maximally legible settings, highlighting practical impact for nuanced human–robot interaction. The approach enables context-dependent expressiveness in robot motion, with potential extensions to real human studies and more complex interaction scenarios.

Abstract

Legibility of robot motion is critical in human-robot interaction, as it allows humans to quickly infer a robot's intended goal. Although traditional trajectory generation methods typically prioritize efficiency, they often fail to make the robot's intentions clear to humans. Meanwhile, existing approaches to legible motion usually produce only a single "most legible" trajectory, overlooking the need to modulate intent expressiveness in different contexts. In this work, we propose a novel motion generation framework that enables controllable legibility across the full spectrum, from highly legible to highly ambiguous motions. We introduce a modeling approach based on an Information Potential Field to assign continuous legibility scores to trajectories, and build upon it with a two-stage diffusion framework that first generates paths at specified legibility levels and then translates them into executable robot actions. Experiments in both 2D and 3D reaching tasks demonstrate that our approach produces diverse and controllable motions with varying degrees of legibility, while achieving performance comparable to SOTA. Code and project page: https://legibility-modulator.github.io.

Figures (3)

• Figure 1: Overview of Legibility Modulator. The Intent-Aware Path Diffuser generates an intermediate path conditioned on the desired intent expressiveness (legibility score). The Path-Guided Diffusion Policy then translates this path into executable robot actions, producing motion that aligns with the specified level of legibility.
• Figure 2: Detailed architecture of our two-stage conditional diffusion framework. (A) The framework takes as input the initial environment state $s_0$, a candidate goal $g$, the IPF that encodes the probabilistic structure of goal inference, and a continuous legibility score$\ell(\xi)$ specifying the desired level of expressiveness. (B) The Intent-Aware Path Diffuser is a U-Net conditioned via FiLM layers on the inputs, which generates a path $\tau$ using diffusion process. (C) The Path-Guided Diffusion Policy is trained on trajectories generated by the QD algorithm, using the corresponding paths as conditioning inputs. The Intent-Aware Path Diffuser is trained on the same QD trajectories, labeled with potential-based legibility scores.
• Figure 3: Results in the 3D Block Reaching (Target Red), showing smooth modulation from ambiguous to legible motions. Similar behaviors are also observed for the alternative target and in the 2D environment.