Year: 2024
Type: conference-paper
Abstract: We present a novel human-aware epistemic planning frameworkdesigned for collaborative human-robot interactions,specially tailored for situations where the agents’ shared executionexperiences can be interrupted by the uncontrollablenature of humans. Our objective is to generate a robot policythat accounts for such uncontrollable behaviors, thus enablingthe anticipation of potential progress achieved by therobot when the experience is not shared, e.g., when humansare briefly absent from the shared environment to complete asubtask. But this anticipation is considered from the perspectiveof humans who keep an estimated robot’s model. As afirst step to address it, we propose a general planning frameworkand build a solver based on AND/OR search which integratesknowledge reasoning; this includes assessing situationsby perspective taking. Our approach dynamically modelsand manages the expansion or contraction of potentialworlds while tracking whether or not agents share the task executionexperiences. This helps the planner (or the robot) toprepare itself with a set of worlds that humans would considerpossible. The robot assesses the situation from the humanperspective and removes the worlds that it has reason to thinkare impossible. However, there might still be an impossibleworld that is indistinguishable from the real world. In differentsituations, thanks to our planning framework, the robot’spolicy built offline can determine an appropriate course of action,such as answering human queries, explicitly communicatingsome fact without being annoying, or taking appropriateaction in the presence of the human to help them narrowdown the possibilities further, facilitating collaboration. Ourpreliminary experiments show that the framework is effectivefor behavior planning in different situations. We discuss thepractical issues in different problem settings.
Source: HAL
Authors: Shekhar, Shashank, Favier, Anthony, Alami, Rachid
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4399026846
