Reproduce “Learning to Move with Affordance Maps” on OpenResearch

About this paper

William Qi, Ravi Teja Mullapudi, Saurabh Gupta, Deva Ramanan

A hybrid navigation architecture learns spatial "affordance maps" via self-supervised interaction, enabling mobile robots to robustly and safely navigate dynamic, semantically complex environments. This method achieved up to 60% greater exploration coverage and 45% higher goal-directed navigation success over geometry-only baselines, demonstrating superior sample efficiency compared to end-to-end reinforcement learning.

Key insights

• Self-supervised learning, leveraging active interaction feedback, provides a sample-efficient way to generate dense pixel-level labels for learning complex navigability affordances, implicitly capturing safety margins around dynamic elements.
• Active trajectory sampling, where the agent seeks out areas of high model uncertainty, accelerates learning of rare or difficult-to-identify semantic features, enhancing overall model performance.
• Decomposition into an explicit, interpretable spatial affordance map allows classical, robust planners to effectively reason about complex semantic and dynamic environmental constraints, bridging the gap between learning and geometry.

Summary by alphaXiv.