Auto-Navigator: Decoupled Neural Architecture Search for Visual Navigation

Tianqi Tang, Xin Yu, Xuanyi Dong, Yi Yang

Abstract

Visual navigation agents typically rely on generic image classification backbones, which are not optimized for spatial reasoning. Auto-Navigator introduces a decoupled Neural Architecture Search (NAS) framework that tailors visual perception networks specifically for navigation. By leveraging imitation learning as a dense supervision signal and auxiliary geometric tasks, Auto-Navigator produces efficient architectures that significantly outperform standard hand-crafted models in both success rate and computational efficiency.

Introduction: The Limitations of Generic Backbones

In visual navigation, an agent must map visual observations to movement actions to reach a goal. Most existing systems utilize "off-the-shelf" visual encoders like ResNet, which were designed for object classification. However, navigation demands an understanding of scene geometry and spatial layout rather than just object identity.

While Neural Architecture Search (NAS) offers a path to automated design, applying it to navigation is challenging. Reinforcement Learning (RL), the standard training paradigm for navigation, provides sparse and "weak" rewards. When the network architecture is in flux during search, RL policies often become unstable, leading to poor convergence.

The Auto-Navigator Framework

Auto-Navigator addresses these challenges through a decoupled, multi-stage training pipeline that stabilizes the search process.

Overview of the Auto-Navigator architecture and search space, showing the multi-task learning setup and cell-level search operations.

The framework consists of four key components:

  1. Auxiliary Task Pre-training: The perception module is pre-trained on geometric tasks, such as depth prediction and surface normal estimation, to ground the search space in spatial awareness.
  2. Imitation Learning (IL) for NAS: Instead of RL, the search phase uses Behavioral Cloning on optimal paths. This provides a dense supervision signal, allowing the NAS to identify architectural choices that directly improve action accuracy.
  3. Decoupled Search Space: The system utilizes a cell-based search space. Two variants were tested: Auto-Navigator-S (Small, 5 operations) and Auto-Navigator-L (Large, 9 operations including depthwise separable convolutions).
  4. RL Fine-tuning: After the optimal architecture is identified via IL, the agent is fine-tuned using Proximal Policy Optimization (PPO) to enhance generalization in environments where expert paths are unavailable.

Experimental Evaluation

The researchers evaluated Auto-Navigator using the Habitat simulator on the Gibson and Matterport3D (MP3D) datasets. The primary task was PointGoal navigation, requiring the agent to reach a specific coordinate relative to its start point.

Key Results and Insights

Auto-Navigator demonstrated significant improvements over baseline models like SplitNet.

Qualitative comparison of navigation paths between SplitNet and Auto-Navigator variants. Auto-Navigator-L achieves higher SPL by following more efficient routes.
  • Superior Performance: On the Gibson validation set, Auto-Navigator-L achieved a Success weighted by Path Length (SPL) of 76.35%, compared to 70.10% for the SplitNet baseline.
  • Efficiency: The framework is remarkably efficient, with training times roughly 250 times faster than state-of-the-art models like DD-PPO, while utilizing significantly fewer parameters.
  • Robustness: The agents showed improved performance in "difficult" episodes involving long distances and complex obstacles. The "Small" variant (0.39M parameters) remained competitive, proving that specialized small architectures can outperform generic larger ones.

Conclusion and Future Work

Auto-Navigator proves that decoupling architecture search from reinforcement learning is a viable strategy for visual navigation. By using imitation learning and geometric priors, the framework designs perception networks that are both more effective and more efficient than hand-crafted alternatives.

Future research may extend this approach to more complex tasks such as ObjectGoal or AreaGoal navigation. Additionally, while the current method optimizes the internal "cells" of the network, fully automating the macro-architecture (the overall stacking of these cells) remains an open challenge.

Summary updated

Abstract

Existing visual navigation approaches leverage classification neural networks to extract global features from visual data for navigation. However, these networks are not originally designed for navigation tasks. Thus, the neural architectures might not be suitable to capture scene con-tents. Fortunately, neural architecture search (NAS) brings a hope to solve this problem. In this paper, we propose an Auto-Navigator to customize a specialized network for visual navigation. However, as navigation tasks mainly rely on reinforcement learning (RL) rewards in training, such weak supervision is insufficiently indicative for NAS to optimize visual perception network. Thus, we introduce imitation learning (IL) with optimal paths to optimize navigation policies while selecting an optimal architecture. As Auto-Navigator can obtain a direct supervision in every step, such guidance greatly facilitates architecture search. In particular, we initialize our Auto-Navigator with a learnable distribution over the search space of visual perception architecture, and then optimize the distribution with IL supervision. Afterwards, we employ an RL reward function to fine-tune our Auto-Navigator to improve the generalization ability of our model. Extensive experiments demonstrate that our Auto-Navigator outperforms baseline methods on Gibson and Matterport3D without significantly increasing network parameters.
BibTeX
@inproceedings{tang2021auto,
  author = {Tianqi Tang and Xin Yu and Xuanyi Dong and Yi Yang},
  title = {{Auto-Navigator: Decoupled Neural Architecture Search for Visual Navigation}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2021},
  doi = {10.1109/wacv48630.2021.00379},
}
Figure