Reproduce “Integrating Deep RL and Bayesian Inference for ObjectNav in Mobile Robotics” on OpenResearch

About this paper

João Castelo-Branco, José Santos-Victor, Alexandre Bernardino

Researchers from Instituto Superior Técnico and Institute For Systems and Robotics developed a hybrid framework for autonomous object search, combining Bayesian inference for uncertainty modeling with deep reinforcement learning for adaptive action selection. This approach attained perfect or near-perfect success rates (100% and 99%) and improved search efficiency in larger indoor environments, reducing actions by 23% and distance by 18% compared to a purely probabilistic method.

Key insights

• Explicitly maintaining and updating a probabilistic belief map over target locations is crucial for achieving high success rates in object search under partial observability.
• A learned Deep Reinforcement Learning (DRL) policy, when informed by a Bayesian belief map, can adaptively select navigation goals to improve search efficiency, particularly in larger environments.
• While structured exploration (e.g., clustering) is beneficial, it benefits from integration with probabilistic reasoning to avoid exploring regions where evidence points elsewhere, especially in complex scenes.

Summary by alphaXiv.