SF-ODNav: Successor Feature Framework for Map-Less Target-Driven Outdoor Visual Navigation
Junzhe Wu, Jiaming Zhang, Tingrong Zhang, Ruining Tao, Huy Tran, Girish Chowdhary
AI summary
Problem
Deep reinforcement learning for visual navigation struggles to generalize in dynamic, unstructured outdoor environments where GPS is unreliable and high-resolution maps are unavailable. Existing methods often require extensive retraining or fail to transfer effectively from simulation to real-world continuous spaces.
Approach
The authors decouple environmental dynamics from task rewards using a successor feature framework, training a goal-conditioned policy on grid-world environments constructed from real outdoor images. Multi-stage encoder pretraining and noise-perturbed successor feature learning enable rapid adaptation to novel environments and real-world deployment.
Key results
- 71.34% and 90.06% overall success rates in outdoor grid-worlds, surpassing DQN and DSFN baselines
- Effective within-domain transfer via successor feature layer retraining with a fixed visual encoder
- Successful real-time deployment from discrete grid-world training to continuous outdoor navigation
- Ablation confirms contrastive/Siamese losses, multi-step learning, and noise perturbation drive efficiency and stability
Why it matters
Provides a scalable, map-less navigation solution for mobile robots operating in GPS-denied or unstructured outdoor settings like agriculture and dense urban areas.
Abstract
Traditional deep reinforcement learning-based vi- sual navigation techniques face challenges in dynamic and unstructured outdoor environments, particularly in the absence of high-resolution maps and GPS signals. This paper presents a deep reinforcement learning-based approach for target-driven visual navigation without explicit localization and mapping in outdoor settings, using the successor feature (SF) frame- work to enhance the model’s transfer learning. This design enables effective knowledge transfer across tasks, allowing the model to adapt to novel environments with zero-shot or few-shot fine-tuning. To facilitate training and evaluation, we design grid-world environments constructed from real-world outdoor images, providing realistic yet controlled conditions for developing and testing deep reinforcement learning-based navigation. Experimental results demonstrate that our method can adapt effectively in outdoor environments, both within the same domain and across different domains. Moreover, despite being trained in a discrete grid-world setting, the model is successfully deployed in real time within the same area, maintaining robust performance and highlighting its strong transferability to continuous, real-world conditions.