SuReNav: Superpixel Graph-based Constraint Relaxation for Navigation in Over-constrained Environments

TL;DR

SuReNav tackles over-constrained navigation in semi-static environments by learning to relax regional constraints in a region-graph representation. It combines a superpixel graph with a GNN-based relaxation cost estimator and a differentiable graph-based A* planner, trained end-to-end from human demonstrations, and operates in an online loop that interleaves planning with execution. Key contributions include formalizing constrained-region relaxation, introducing GraphGPS-based relaxation costs, and demonstrating human-like, safe, and efficient navigation on 2D maps, OpenStreetMap 3D maps, and a real quadruped robot. The approach generalizes beyond fixed regional costs and achieves high human-likeness while maintaining completeness and safety in dynamic semi-static scenarios, offering practical impact for autonomous navigation in urban environments.

Abstract

We address the over-constrained planning problem in semi-static environments. The planning objective is to find a best-effort solution that avoids all hard constraint regions while minimally traversing the least risky areas. Conventional methods often rely on pre-defined area costs, limiting generalizations. Further, the spatial continuity of navigation spaces makes it difficult to identify regions that are passable without overestimation. To overcome these challenges, we propose SuReNav, a superpixel graph-based constraint relaxation and navigation method that imitates human-like safe and efficient navigation. Our framework consists of three components: 1) superpixel graph map generation with regional constraints, 2) regional-constraint relaxation using graph neural network trained on human demonstrations for safe and efficient navigation, and 3) interleaving relaxation, planning, and execution for complete navigation. We evaluate our method against state-of-the-art baselines on 2D semantic maps and 3D maps from OpenStreetMap, achieving the highest human-likeness score of complete navigation while maintaining a balanced trade-off between efficiency and safety. We finally demonstrate its scalability and generalization performance in real-world urban navigation with a quadruped robot, Spot.

Figures (6)

• Figure 1: A capture of regional-constraint relaxation experiment in an over-constrained navigation environment. A quadruped robot pre-plans a path on a known map, and when the path is blocked, our method finds a new path by relaxing a compact area with the least risky constraints to enable safe and efficient human-like navigation.
• Figure 2: Overview architecture of SuReNav that automatically relaxes regional constraints while planning a graph path $\mathcal{X}_G$ in semi-static navigation environment. (a) In the training phase, SuReNav generates a superpixel graph from a 2-dimensional map with safety-relevant semantic features. A neural relaxation cost estimator $\Psi_G$ then computes a node-wise relaxation cost, which is incorporated as an optimization term in a differentiable search process. This enables end-to-end training of the model from human demonstrations. (b) During the planning phase, the estimator generates node-wise relaxation costs, and a discrete graph search process selects regions to relax, inducing an admissible space to facilitate path planning.
• Figure 3: Real-world evaluation in semi-static settings. The quadruped robot begins by following a path planned using an outdated prior map. Upon encountering newly constrained regions, SuReNav dynamically relaxes a red-colored soft-constraint region and adapts its behavior to perform human-like navigation that balances safety and efficiency. Video demonstrations are available in the supplementary material.
• Figure 4: Comparison of the proposed SuReNav and four baseline method in $300$ simulated navigation environments from three urban areas— Milan, Baltimore, and Capitol Hill. We evaluate average SPL and Total Risk metrics where performance improves toward the lower-right corner that is higher SPL and lower Total Risk.
• Figure 5: Examples of navigation behaviors across three urban scenarios. Green circles and yellow stars denote start and goal locations, respectively. The ivory, green, light orange, gray, orange, and blue regions correspond to sidewalks, lawns, crosswalks, roads, buildings, and water areas, respectively. Colored line segments indicate the trajectories of individual navigation examples.