Two-Stage Depth Enhanced Learning with Obstacle Map For Object Navigation

2024-06-20

arXiv.org

AI Summary

Existing visual object navigation methods suffer from two key limitations: (1) coupling search and path planning into a single stage with shared reward signals, leading to insufficient training or overfitting; and (2) relying on generic visual encoders that ignore depth and dynamic obstacle information, hindering effective policy learning. To address these issues, we propose a decoupled two-stage navigation framework: (i) a differentiated reward mechanism that separately optimizes target search coverage and path navigation accuracy; (ii) an RGB-D pre-trained depth-aware feature extractor integrated with an online-constructed obstacle map and semantic cues for multimodal state representation; and (iii) end-to-end joint optimization via two-stage reinforcement learning. Evaluated on AI2-Thor and RoboTHOR, our method achieves significant improvements over state-of-the-art approaches—higher success rates, improved path efficiency, and substantially reduced collision and deadlock rates.

Technology Category

• Search and Optimization: Learning to Search
• Computer Vision: Visual Reasoning & Symbolic Representations
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

Abstract

The task that requires an agent to navigate to a given object through only visual observation is called visual object navigation (VON). The main bottlenecks of VON are strategies exploration and prior knowledge exploitation. Traditional strategies exploration ignores the differences of searching and navigating stages, using the same reward in two stages, which reduces navigation performance and training efficiency. Our study enables the agent to explore larger area in searching stage and seek the optimal path in navigating stage, improving the success rate of navigation. Traditional prior knowledge exploitation focused on learning and utilizing object association, which ignored the depth and obstacle information in the environment. This paper uses the RGB and depth information of the training scene to pretrain the feature extractor, which improves navigation efficiency. The obstacle information is memorized by the agent during the navigation, reducing the probability of collision and deadlock. Depth, obstacle and other prior knowledge are concatenated and input into the policy network, and navigation actions are output under the training of two-stage rewards. We evaluated our method on AI2-Thor and RoboTHOR and demonstrated that it significantly outperforms state-of-the-art (SOTA) methods on success rate and navigation efficiency.

Problem

• Decoupling search and pathfinding in object navigation tasks
• Improving spatial perception with depth-enhanced visual encoding
• Assessing search efficiency with a new weighted metric

Innovation

• Two-Stage Reward Mechanism decouples searching and pathfinding
• Depth Enhanced Masked Autoencoders improve spatial perception
• New metric SSSPL evaluates searching ability and efficiency