Embedded Files
Jose Lavariega-Gomez, Nicola Irmiger, Erica Tevere, Patrick Spieler
Abstract:
Autonomous long-range navigation in unstructured environments is challenging due to limited sensing range, partial observability, and the difficulty of estimating the long- term utility of unexplored regions. We present COMPASS, a visual-geometric navigation framework that learns a goal- conditioned frontier utility policy over graph-structured spatial memory and semantic visual observations. COMPASS combines a Graph Convolutional Network (GCN) operating over a persistent navigation graph with DINOv3 visual embeddings extracted from map frontier observations. The graph captures global topology and traversability structure, while visual embeddings provide non-local cues regarding paths, openings, and terrain continuity. A Deep Q-Learning policy estimates frontier utility under delayed rewards, enabling behaviors such as backtracking and dead-end recovery. To train the policy efficiently, we introduce a lightweight graph-based rollout simulator that expands limited field data into millions of navigation experiences. Experimental results in unseen outdoor environments demonstrate improved long-range navigation performance compared to heuristic frontier-selection baselines.
In our problem setting, an autonomous legged platform must reach a goal specified in relative coordinates through an unknown, unmapped environment. The robot must do so by selecting frontiers: the nodes at the limit of mapped and unmapped space. Each frontier selection allows the local construction of a navigation graph, and unlocks more frontiers, dead-ends or closes a loop.
Our research objective is to find if visual foundation models and graph neural networks can enhance the features of the frontiers, in order to efficiently navigate to the goal. Notably, we do so without using engineered heuristics on the visual domain.
Research Goal: Can foundation models and geometric planning achieve efficient Field Navigation without needing engineered heuristics?
Approach: A Learned Frontier Scorer that considers Visual and low-detail graph features to score and select the frontier that minimizes time to goal.
Compass: Using VFMs and GNNs for Long Range Navigation.
Compass: Using VFMs and GNNs for Long Range Navigation.
We use three main feature modules: A DINOv3 model processes images of the active frontiers to extract mean and standard deviation across the tokens. A Graph Convolutional Network processes the current graph memory and produces a latent that scores frontiers conditioned on the current goal location, outside of the graph. The Frontier Scorer takes both of these latent spaces and produces a final score across all frontiers. The Frontier Scorer is trained in a navigation-level simulator, using only cached images from data capture, and the graph structure.
Our Frontier Scoring module takes as inputs:
The constructed navigation graph at the current step.
Frontier locations within the navigation graph.
Their alignment, estimated progress, and step size towards the goal.
Images of the current frontiers, as seen from within the navigation graph.
And it outputs a goal-conditioned frontier score, to select the next frontier to navigate to.
Simulation Results:
Simulation test where agent must reach the goal within 50 steps. (Reconstructed frontier graph visualized at steps 0, 15, 30, 40).
The Baseline gets locally stuck by selecting geometrically close frontiers but unable to backtrack in time, as it cannot recognize a dead-end.
The GCN Prior is able to backtrack about multiple dead-ends and reaches the goal.
Compass backtracks from these dead-ends earlier, and is reaches the goal in the least amount of steps.
Deployment Results:
We deployed on a mixed urban-forested outdoor environment, with diversity across buildings and obstacle types.
We compare Compass to a Navigation Graph Planner used for the DARPA SubT Challenge, based entirely on geometric features.
In our first deployment, we observe that Compass's usage of visual features allows it to identify a potential dead end through a building, and chooses a more efficient path towards the goal. The Navigation Graph Planner does not recognize these path inefficiencies and searches through dead-ends and ultimately backtracks.
In the second deployment, a large negative obstacle impedes passage and is not visible until close by. The Navigation Graph Planner avoids the obstacle at the last minute, whereas Compass recognizes an opening through visual features ~30m earlier.
The work was presented at the International Conference of Robotics and Automation (ICRA) 2026 in Vienna, Austria.
It was part of two workshops:
1st Workshop on Long-Term Deployments in the Wild (LoWi), June 1
Workshop on Field Robotics, June 5
Presented as part of the Poster Session in both.
Poster BibTeX Citation:
@inproceedings{lowi2026_p10,
author = {Lavariega-Gomez, Jose and Irmiger, Nicola and Tevere, Erica and Spieler, Patrick},
title = {COMPASS: Learning Global Spatial Context for Long-Range Robot Navigation},
booktitle = {IEEE International Conference on Robotics and Automation (ICRA) Workshop (LoWi 2026)},
year = {2026},
note = {Non-archival Workshop Paper},
url = {https://mobile-robotics-hub.github.io/workshop2026/papers/LoWi2026_P10.pdf}
}
000_Thesis_FinalPresentation.pptxThesis Presentation
(Manuscript Coming Soon)
Page updated
Google Sites
Report abuse