Beyond Memory Limits

Project Overview

Project Demo

FLAVOR couples a fixed-size local map with historical visibility, persistent frontier tasks, and sparse topology to plan an exploration tour and executable trajectory. — **Local geometry meets mission-scale memory.** FLAVOR preserves observation evidence, unfinished tasks, and return routes while keeping dense geometry within a robot-centric map.

Research overview

Abstract

Autonomous UAV exploration supports large-scale surveying and inspection under strict onboard memory and computation constraints. Retaining globally queryable geometry makes memory and query costs grow with the mapped workspace. A robot-centric local map bounds this geometric state, but its finite window discards the observation and task context needed for environment-wide exploration.

We present FLAVOR, a hierarchical spatio-temporal framework that confines dense metric geometry to a local map while retaining compact visibility evidence, unfinished frontier tasks, odometry anchors, and topological routes. A cascaded frontier mechanism preserves unresolved objectives as the map moves. When the UAV returns to a task region, current observations reconstruct its frontier and viewpoint, and the recovered route is checked and repaired before trajectory generation.

Experiments in large-scale simulated environments and two real-world deployments demonstrate efficient exploration beyond the local-map extent with substantially reduced map-representation memory. Dense-map capacity remains independent of explored range; exploration-dependent growth is confined to compact observation, task, and route context.

The FLAVOR framework

Remember where. Reobserve how.

Retained mission context guides where exploration resumes. Current local geometry determines how each task is executed.

01 / REPRESENTATION

Bounded local geometry

A fixed-size robot-centric map supports occupancy queries, frontier validation, collision checking, and trajectory generation.

02 / TASK MEMORY

Persistent frontier tasks

LiDAR-keyframe visibility and cascaded frontier updates preserve observation evidence and unfinished objectives across map shifts.

03 / RETURN PLANNING

Anchor-associated planning

Sparse topological routes guide the UAV back to pending tasks, which are reconstructed and validated using current observations.

FLAVOR workflow: local mapping, keyframe visibility, frontier updates and a spatio-temporal graph feed global task planning, local route validation, safe corridor construction and trajectory optimization. — The closed-loop pipeline links hierarchical mapping and task planning to route validation, safe corridor construction, and trajectory optimization.

In simulation

Simulation Experiments

Evaluation in large-scale point-cloud environments with MARSIM as the closed-loop LiDAR and vehicle-simulation backend. All planners use a maximum planning speed of 4.0 m/s.

Cave

SIM 01

224 × 330 × 41 m

Interconnected tunnels, narrow passages, and vertical drops in the DARPA SubT Cave World scene.

Forest

SIM 02

108 × 88 × 5 m

A vegetation-rich forest reconstructed from terrestrial laser scans collected in Guangxi, China.

Lava Tube

SIM 03

313 × 62 × 25 m

A long, confined passage with irregular geometry in a lava-tube scene in Hainan, China.

Exploration beyond the memory window

Across the three benchmark scenes, FLAVOR completes exploration with 31.98–41.88% less time and 94–99% less map-representation memory than EPIC.

FLAVOR results in the three simulated scenes. Reductions are relative to EPIC.
Scene	Completion time	Time reduction	Map memory	Memory reduction
Cave	845.20 s	31.98%	41.93 MB	97.22%
Forest	276.87 s	41.88%	13.48 MB	99.00%
Lava Tube	152.30 s	41.26%	14.56 MB	94.00%

Onboard autonomy

Real-world Experiments

The complete exploration pipeline runs onboard an Intel N305 computer with a Livox MID360 LiDAR. Both deployments extend beyond the local-map width, with a maximum flight speed of 1.2 m/s.

Forest

FIELD 01

60 × 27 × 3 m exploration region

Autonomous exploration through unstructured vegetation, using retained visibility evidence and frontier tasks to maintain exploration continuity.

Complex Construction Site

FIELD 02

50 × 27 × 3.5 m exploration region

Exploration amid structural clutter and transient dust returns, with occupancy and collision queries confined to the sliding local map.

Acknowledgements

We thank the developers of MARSIM for the UAV simulation platform, ROG-Map for its efficient robot-centric map representation, and EPIC for making their exploration framework openly available. We gratefully acknowledge their contributions to the open-source robotics community.

BibTeX

@misc{zhang2026memorylimitshierarchicalspatiotemporal,
      title={Beyond Memory Limits: A Hierarchical Spatio-Temporal Map Representation for Large-Scale 3D Autonomous Exploration}, 
      author={Yizhe Zhang and Jianping Li and Yilun Gao and Liangliang Yin and Pengfei Wan and Bisheng Yang},
      year={2026},
      eprint={2611.11489},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2611.11489}
}