IntentNav: Learning Spatial-Visual Object Navigation from Human Demonstrations
Abstract Overview
IntentNav is a framework for ObjectNav that learns human-like search policies from human demonstrations rather than relying on low-level action imitation alone. The method converts low-level human trajectories into candidate-level supervision with Frontier-based Human-Intent Labeling, which infers which frontier best matches the demonstrator’s future search direction. It represents navigation decisions in a shared bird’s-eye-view candidate space that combines persistent spatial memory, candidate-specific egocentric visual memory, and agent-centric geometry. A vision-language model is then trained to select among grounded frontier or target candidates using an Intent-Aligned Objective designed to preserve directional consistency during exploration.
Novelty
The paper’s main novelty is a candidate-level spatial-visual imitation framework that infers high-level search intent from low-level human demonstrations through frontier evolution. It also introduces a unified BEV candidate interface that grounds both exploration frontiers and target commitments with paired visual memory and geometry, rather than predicting free-form coordinates or low-level actions.
Results
IntentNav reports state-of-the-art performance among learning-based methods on MP3D, HM3D-v1, and HM3D-v2 ObjectNav benchmarks, achieving 53.8/23.1 SR/SPL on MP3D, 70.5/34.6 on HM3D-v1, and 82.2/38.5 on HM3D-v2. Ablations indicate that candidate visual memory, agent-centric geometry, reserved candidate-ID selection, and the Intent-Aligned Objective each contribute to stronger performance and more directed search behavior. The same MP3D-trained checkpoint is also deployed zero-shot on wheeled, quadruped, and humanoid robots, with approximately 0.8 s per high-level policy inference.
Key Points
- Frontier-based Human-Intent Labeling turns low-level human demonstrations into candidate-level supervision by reasoning over future frontier evolution.
- IntentNav’s policy operates in a BEV candidate space that jointly uses spatial memory, candidate-aligned egocentric visual evidence, and explicit relative geometry.
- The method shows strong benchmark generalization and real-world transfer, with reported state-of-the-art learning-based results on three ObjectNav benchmarks and zero-shot deployment across multiple robot embodiments.