Uncertainty-Calibrated System-2 MLLM Deliberation and System-1 VLA Control with Selective Abstention for Autonomous Earth Observation

Autonomous Earth observation requires reliable aerial navigation under environmental degradation and distribution shift. We propose U-S2A-VLA, an uncertainty-calibrated dual-system framework that combines deliberative planning, reactive control, and selective abstention. Its asynchronous spatiotemporal context engine maintains a causal graph and issues subgoals at 1 Hz, while an edge policy uses abnormal-aware distillation to control flight at 15 Hz. A conformal uncertainty gate separates aleatoric and epistemic uncertainty to guide execution, safe hovering, or re-scouting. Experiments across AerialVLN, FloodNet, and RS-Agent show a 47.85% navigation success rate, 4.25 percentage points above the previous state of the art, and a 3.45% expected calibration error, with 18.4 ms edge execution latency.

Authors

Publication Details

Published
2026-11-13
DOI
https://doi.org/10.22541/essoar.15012148/v1
Primary Topic
AI-based Problem Solving and Planning
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Uncertainty-Calibrated System-2 MLLM Deliberation and System-1 VLA Control with Selective Abstention for Autonomous Earth Observation

Haoran Lian, Ziyang Pei, Siyuan Ke, Yuxun Cen et al.
AI-based Problem Solving and Planning
preprint

Uncertainty-Calibrated System-2 MLLM Deliberation and System-1 VLA Control with Selective Abstention for Autonomous Earth Observation

Haoran Lian, Ziyang Pei, Siyuan Ke, Yuxun Cen, Zhenyu Qiao
preprint en

Abstract

Autonomous Earth observation requires reliable aerial navigation under environmental degradation and distribution shift. We propose U-S2A-VLA, an uncertainty-calibrated dual-system framework that combines deliberative planning, reactive control, and selective abstention. Its asynchronous spatiotemporal context engine maintains a causal graph and issues subgoals at 1 Hz, while an edge policy uses abnormal-aware distillation to control flight at 15 Hz. A conformal uncertainty gate separates aleatoric and epistemic uncertainty to guide execution, safe hovering, or re-scouting. Experiments across AerialVLN, FloodNet, and RS-Agent show a 47.85% navigation success rate, 4.25 percentage points above the previous state of the art, and a 3.45% expected calibration error, with 18.4 ms edge execution latency.

AI-based Problem Solving and Planning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon