[Paper Review] Learning to Map for Active Semantic Goal Navigation

Georgios Georgakis, Bernadette Bucher+3 more|arXiv (Cornell University)|Jun 29, 2021

Multimodal Machine Learning Applications58 references4 citations

TL;DR

This paper proposes Learning to Map (L2M), a novel framework for active semantic goal navigation that generates semantic maps beyond the agent's field of view and uses prediction uncertainty to guide long-term goal selection. By actively training an ensemble of semantic segmentation models using epistemic uncertainty as an information gain objective, L2M improves success rates on object-goal navigation in unseen environments, achieving 80% success with optimal components in the Matterport3D benchmark.

ABSTRACT

We consider the problem of object goal navigation in unseen environments. Solving this problem requires learning of contextual semantic priors, a challenging endeavour given the spatial and semantic variability of indoor environments. Current methods learn to implicitly encode these priors through goal-oriented navigation policy functions operating on spatial representations that are limited to the agent's observable areas. In this work, we propose a novel framework that actively learns to generate semantic maps outside the field of view of the agent and leverages the uncertainty over the semantic classes in the unobserved areas to decide on long term goals. We demonstrate that through this spatial prediction strategy, we are able to learn semantic priors in scenes that can be leveraged in unknown environments. Additionally, we show how different objectives can be defined by balancing exploration with exploitation during searching for semantic targets. Our method is validated in the visually realistic environments of the Matterport3D dataset and show improved results on object goal navigation over competitive baselines.

Motivation & Objective

• To address the challenge of navigating to semantic targets in unseen indoor environments where spatial and semantic variability hinders generalization.
• To overcome the limitation of existing methods that rely only on observed regions by enabling agents to predict and reason about unobserved semantic content.
• To develop an active learning strategy that uses model uncertainty to guide data sampling and improve semantic map generalization.
• To balance exploration and exploitation in goal selection using uncertainty-based planning objectives such as upper confidence bounds.
• To demonstrate that semantic priors learned through spatial prediction can generalize to unknown environments and improve navigation performance.

Proposed method

• The framework employs an ensemble of two-stage semantic segmentation models trained on top-down maps to predict both occupancy and semantic regions in unobserved areas.
• Model uncertainty is estimated via disagreement among ensemble members, following methods from Pathak et al. (2019) and Seung et al. (1992), to quantify confidence in predictions.
• An active learning strategy selects training samples based on information gain, with epistemic uncertainty emerging as the most effective objective for improving map quality.
• Long-term navigation goals are selected using uncertainty-aware planners, including upper confidence bound (UCB) strategies that balance exploration and exploitation.
• The system integrates a local policy for navigation and evaluates performance using success rate and success rate with oracle stop/goal conditions.
• The method is evaluated in the Habitat simulator using the Matterport3D dataset, with ablation studies on components like local policy and stop criteria.

Experimental results

Research questions

• RQ1 Can active learning using model uncertainty improve the generalization of semantic map predictions in unobserved regions for navigation?
• RQ2 How does uncertainty-based goal selection compare to random or greedy strategies in balancing exploration and exploitation during semantic navigation?
• RQ3 To what extent does learning semantic priors through spatial hallucination improve success rates in unseen environments?
• RQ4 What is the contribution of the local policy and stop decision mechanism to overall navigation performance?
• RQ5 How do different information gain objectives during training affect the quality of semantic map predictions and downstream navigation success?

Key findings

• The L2M framework achieves a mean success rate of 80% on a challenging subset of test episodes when combining the best-performing components: GtPath for local navigation and OracleStop for goal detection.
• Epistemic uncertainty was found to be the most effective information gain objective for active data sampling, significantly improving semantic map quality and navigation performance.
• Using uncertainty-based planning with upper confidence bounds led to better balance between exploration and exploitation, resulting in higher success rates than baseline strategies.
• The ablation study revealed that the local policy is a major bottleneck, as replacing it with shortest-path estimates (GtPath) substantially improved performance.
• The model's ability to predict semantic content beyond the field of view enables generalization to unseen environments, outperforming competitive baselines on the Matterport3D dataset.