ObjectReact: Learning Object-Relative Control for Visual Navigation

Sourav Garg    Dustin Craggs    Vineeth Bhat¹    Lachlan Mares    Stefan Podgorski    Madhava Krishna¹    Feras Dayoub    Ian Reid   

¹ Robotics Research Center, IIIT Hyderabad, India   

ObjectReact introduces a learning-based framework for **object-relative control** in visual navigation. Instead of learning control policies in absolute coordinate spaces, ObjectReact grounds actions relative to objects in the scene, enabling more robust and semantically aware navigation. The approach leverages learned object-centric representations to improve generalization and adaptability across environments. This formulation helps agents navigate effectively by reasoning about their actions with respect to landmarks and objects rather than fixed poses alone. The method demonstrates effectiveness on challenging navigation benchmarks, highlighting its ability to support long-horizon, object-conditioned tasks. *Presented at the 9th Conference on Robot Learning (CoRL 2025), Seoul, Korea.*