NavMoE: Hybrid Model and Learning-Based Traversability Estimation for Local Navigation Via Mixture of Experts
Botao He, Amir Hossein Shahidzadeh, Yu Chen, Jiayi Wu, Tianrui Guan, Guofei Chen, Dinesh Manocha, Howie Choset, Glen Chou, Cornelia Fermüller, Yiannis Aloimonos
AI summary
Problem
Traversability estimation for robot navigation struggles to balance reliable, robust predictions with efficient computation while accurately encoding both geometric and semantic information across diverse, unseen environments.
Approach
NAVMOE uses a hierarchical Mixture of Experts architecture that routes sensor inputs through domain and terrain routers to selectively activate specialized model-based or learning-based experts, guided by a training-free lazy gating mechanism that prunes unnecessary computations.
Key results
- 35.3% higher average accuracy in traversability map estimation compared to individual experts
- 81.2% average computational cost reduction via lazy gating with less than 2% path quality loss
- Hierarchical routing with domain and pixel-wise terrain specialization for adaptive expert selection
- Two-stage training strategy enabling effective learning with limited high-quality multi-sensory data
Why it matters
Provides a scalable, efficient framework for safe robot navigation that bridges the gap between interpretable geometric models and data-driven learning, benefiting robotics researchers and autonomous system developers.
Abstract
This paper explores traversability estimation for robot navigation. A key bottleneck in traversability estimation lies in efficiently achieving reliable and robust predictions while accurately encoding both geometric and semantic information across diverse environments. We introduce Navigation via Mixture of Experts (NAVMOE), a hierarchical and modular approach for traversability estimation and local navigation. NAVMOE combines multiple specialized models for specific terrain types, each of which can be either a classical model- based or a learning-based approach that predicts traversability for specific terrain types. NAVMOE dynamically weights the contributions of different models based on the input envi- ronment through a gating network. Overall, our approach offers three advantages: First, NAVMOE enables traversability estimation to adaptively leverage specialized approaches for different terrains, which enhances generalization across diverse and unseen environments. Second, our approach significantly improves efficiency with negligible cost of solution quality by introducing a training-free lazy gating mechanism, which is designed to minimize the number of activated experts dur- ing inference. Third, our approach uses a two-stage training strategy that enables the training for the gating networks within the hybrid MoE method that contains nondifferentiable modules. Extensive experiments show that NAVMOE delivers a better efficiency and performance balance than any individual expert or full ensemble across domains, improving cross-domain generalization and reducing average computational cost by 81.2% via lazy gating, with less than a 2% loss in path quality.