Personalized Embodied Navigation for Portable Object Finding

arXiv cs.RO / 4/22/2026

💬 OpinionDeveloper Stack & InfrastructureModels & Research

共有:

Key Points

The paper studies embodied navigation in dynamic settings where the target is a non-stationary portable object moved by human activity rather than a fixed landmark.
It formalizes portable object finding as a personalized habit learning problem and proposes two Transit-Aware Planning (TAP) methods that incorporate target object path information into navigation policies.
TAP improves performance by encouraging the agent to synchronize its travel route with the typical target routes learned from transit patterns.
Evaluations on Dynamic Object Maps (DOMs) and both simulation and real-world tests show sizable gains, including a 21.1% success improvement in MP3D simulation for non-stationary targets and an 18.3% average improvement across real-world transit scenarios.
The authors also describe a real-to-sim pipeline to let other researchers generate TAP-relevant simulations from their own physical environments, supporting further research in this area.

Abstract

Embodied navigation methods commonly operate in static environments with stationary objects. In this work, we present approaches for tackling navigation in dynamic scenarios with non-stationary targets. In an indoor environment, we assume that these objects are everyday portable items moved by human intervention. We therefore formalize the problem as a personalized habit learning problem. To learn these habits, we introduce two Transit-Aware Planning (TAP) approaches that enrich embodied navigation policies with object path information. TAP improves performance in portable object finding by rewarding agents that learn to synchronize their routes with target routes. TAPs are evaluated on Dynamic Object Maps (DOMs), a dynamic variant of node-attributed topological graphs with structured object transitions. DOMs mimic human habits to simulate realistic object routes on a graph. We test TAP agents both in simulation as well as the real-world. In the MP3D simulator, TAP improves the success of a vanilla agent by 21.1% in finding non-stationary targets, while also generalizing better from static environments by 44.5% when measured by Relative Change in Success. In the real-world, we note a similar 18.3% increase on average, in multiple transit scenarios. We present qualitative inferences of TAP-agents deployed in the real world, showing them to be especially better at providing personalized assistance by finding targets in positions that they are usually not expected to be in (a toothbrush in a workspace). We also provide details of our real-to-sim pipeline, which allows researchers to generate simulations of their own physical environments for TAP, aiming to foster research in this area.