Accepted to IROS 2026

AUG: Perception-Aware UAV Graph-Based Planning for GNSS-Denied Navigation in Feature-Sparse Environments

Long-range UAV planning that reasons about geometry, observability, and yaw together.

Enguang Feng, Haohuan Yuan, Hong Zhang, Hao Xu, Yikun Dou, Jiarong Lin, and Xiwang Dong

Corresponding authors: Hao Xu and Jiarong Lin

Beihang University and Nanjing University

AUG plans around estimator health, not only collision avoidance. It incrementally builds a free-space topological graph, runs local MINCO optimization on a robocentric ROG-map, and adjusts yaw so LiDAR keeps observing structure-rich regions during GNSS-denied flight.

Paper PDF Video Citation

Figure 1. Direct routes can be geometrically valid but estimator-hostile. AUG prefers the structure-rich corridor, while baseline planners drift into feature-sparse space and fail.

100% simulation success across 3 sparse scenes

1-2% relative pose error in simulation

0.75-0.81 GB stable memory footprint during planning

2.24% mean GNSS-referenced outdoor drift

Teaser

The video shows route selection, local replanning, and yaw control working as one stack.

It combines the long-horizon graph search with online execution so the vehicle stays in informative geometry instead of merely staying collision-free.

Abstract

When localization weakens, a path that is collision-free can still be mission-failure-prone.

Long-range UAV navigation in GNSS-denied, feature-sparse outdoor environments is difficult because LiDAR-inertial odometry can lose conditioning even when a valid geometric path exists. The problem is not only obstacle avoidance; it is keeping the estimator healthy enough to finish the mission.

AUG addresses this by coupling topology-aware global routing with estimator-aware local optimization. The planner incrementally constructs a free-space graph, performs local trajectory refinement on a robocentric map, and explicitly optimizes yaw to preserve informative sensor observations.

Contributions

Three pieces make the planner practical at long range.

Incremental topology graph

Builds a free-space graph online and prefers corridors that are both traversable and informative, rather than optimizing geometry alone.

Robocentric local planning

Uses a sliding ROG-map and MINCO optimization to keep memory bounded while still generating smooth, real-time trajectories.

Perception-constrained yaw

Scores yaw choices by the quality of visible structure so the LiDAR keeps useful features in view during flight.

Method

Global route, local trajectory, and heading are planned together.

AUG first expands a connectivity graph from local free-space bubbles and feature support. The global route then favors paths through structure-rich corridors, which reduces the chance of entering estimator-degrading regions.

The selected route feeds a local optimizer operating on a sliding robocentric map. In parallel, the yaw module searches for headings that maintain informative LiDAR geometry, making the flight stack behave as a single perception-aware system rather than disconnected modules.

Topology-guided routing

Extracts usable corridors from online observations and avoids weakly observed space when better alternatives exist.

ROG-map trajectory optimization

Optimizes smooth local motion with bounded memory, which is crucial for kilometer-scale missions.

Yaw search and refinement

Adjusts the viewing direction to preserve feature-rich LiDAR returns and suppress estimator drift.

Topology node generation from LiDAR observations — The planner extracts connectivity from live sensing instead of relying on a global prebuilt map.

Results

Reliability improves without a large memory or runtime penalty.

100% success across all 3 simulation scenes

1-2% relative pose error in simulation

24-26 ms replanning time per iteration

2.24% mean GNSS-referenced outdoor drift

Simulation comparison among planners — Baseline planners enter feature-sparse space and fail, while AUG stays on a structure-rich route and completes the task.

EGO feature and error comparison — EGO loses informative structure and error rises quickly once the route leaves the corridor.

SUPER feature and error comparison — SUPER is more scalable, but sparse geometry still drives error upward.

Localization quality comparison across planners — AUG holds low estimator error while the baselines degrade more quickly in the same flight window.

Simulation summary

Scene	Success	ATE	RPE
Scene 1	100%	2.38 m	1.1%
Scene 2	100%	3.68 m	1.5%
Scene 3	100%	3.13 m	1.7%

Experiments

Indoor validation and outdoor flights show the same pattern.

Indoor UAV experiment with LiDAR map inset

Indoor validation

Ten indoor flights of about 30 m confirm stable onboard execution against ground truth.

Outdoor field flight

Ten outdoor flights of about 200 m complete with a mean GNSS-referenced drift ratio of 2.24%.

EGO estimator feature and error over time — EGO keeps seeing features early, then loses the informative structure and accumulates error.

SUPER estimator feature and error over time — SUPER behaves better than EGO, but still degrades in feature-poor corridors.

Topology-guided estimator comparison over time — Topology-guided routing keeps the vehicle in better conditioned regions.

Yaw optimization estimator comparison over time — Yaw optimization is the missing piece for retaining features across the whole flight.

Method	GNSS-referenced drift	Memory	Status
AUG	2.24%	0.92 GB	Success
SUPER	10.13%	0.61 GB	Failure
EGO-Planner	16.80%	2.46 GB	Failure

Citation

Reference this work.

@inproceedings{feng2026aug,
  title = {AUG: Perception-Aware UAV Graph-Based Planning for GNSS-Denied Navigation in Feature-Sparse Environments},
  author = {Feng, Enguang and Yuan, Haohuan and Zhang, Hong and Xu, Hao and Dou, Yikun and Lin, Jiarong and Dong, Xiwang},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year = {2026}
}