DynCur-Geo: Dynamic Curiosity Reward Shaping for Multimodal Active Geo-Localization

Yiming Sun; Yang Zhang; Pengfei Zhu

cs.CVAug 19, 2026

DynCur-Geo: Dynamic Curiosity Reward Shaping for Multimodal Active Geo-Localization

Authors: Yiming Sun, Yang Zhang, Pengfei Zhu

Organizations: School of Automation, Southeast University, Nanjing 210096, China

Abstract

Active geo-localization enables low-altitude UAVs to search for specified targets from limited local aerial observations, supporting time-sensitive applications such as search and rescue and emergency inspection. However, multimodal target cues, restricted views, and sparse feedback make it difficult to balance exploration with target convergence. Existing curiosity-driven methods assign a fixed intrinsic-reward weight throughout search, which can continue rewarding novelty after the agent nears the target and induce detours. We propose DynCur-Geo, a dynamic curiosity framework that adjusts prediction-error intrinsic reward according to remaining target distance. A distance-aware gate encourages early exploration and shifts the policy toward goal-directed behavior near the target, while potential-based reward shaping supplies dense progress guidance. Experiments across multimodal, cross-scene, disaster-affected, and long-range settings show consistent gains over active geo-localization baselines.

Figures & tables

Figure 1: Motivation for distance-gated curiosity. GeoExplorer uses a fixed curiosity weight that can weaken late-stage convergence by rewarding novelty throughout search (orange). Our dynamic gate weights intrinsic reward more strongly far from the target and less near it, promoting exploration followed by goal-directed convergence (blue–green).

Figure 2: Training pipeline and reward construction. Target cues and local aerial observations form a trajectory history processed by a frozen GeoExplorer dynamics model. Prediction error provides intrinsic reward, whose weight is reduced near the target and combined with external progress and PBRS rewards for PPO training.

Table 3

Figure 3: Qualitative visualization of multimodal active geo-localization and test-time policy trajectories. In (a), blue, orange, and green target borders and their corresponding paths denote aerial-image, ground-view-image, and text cues, respectively. In (b), path colors identify DynCur-Geo (blue), GOMAA-Geo (orange), and GeoExplorer (green). Blue squares and yellow circles mark the start and goal cells.

Figure 4: Qualitative xBD cross-disaster visualization. Yellow boxes indicate pre-disaster target regions and red boxes indicate the corresponding post-disaster regions. The lower panels show the pre-disaster target cue and the post-disaster search scene with a test-time policy trajectory; blue squares and yellow circles mark the start and goal cells.

Method	MASA	MM-GAG	xBD-pre	xBD-dis.	Avg.
Random	0.1000	0.0868	0.0925	0.0925	0.0929
PPO	0.2280	0.1991	0.1712	0.1676	0.1915
DiT	0.2280	0.2255	0.2409	0.2365	0.2327
GOMAA-Geo	0.5800	0.5464	0.5612	0.5534	0.5602
GeoExplorer	0.5640	0.5421	0.5233	0.5372	0.5417
DynCur-Geo	0.6200	0.6357	0.5973	0.5984	0.6129

Table 5: Unified

B=12

SR on MASA/A, MM-GAG/A, xBD-pre/A, and xBD-disaster/A; Avg. denotes their macro-average.

Ext.	Int.	Gate	PBRS	$C=4$	$C=5$	$C=6$	$C=7$	$C=8$
$\surd$				0.3617	0.4326	0.5801	0.7305	0.7773
$\surd$			$\surd$	0.3248	0.3546	0.6085	0.8014	0.8823
	$\surd$	1		0.1518	0.1064	0.0794	0.0752	0.0468
$\surd$	$\surd$	1		0.2525	0.3418	0.5816	0.8184	0.8369
$\surd$	$\surd$	Const		0.3787	0.3972	0.5220	0.5929	0.5787
$\surd$	$\surd$	Const	$\surd$	0.3050	0.4099	0.5986	0.7816	0.8426

Table 7: Reward ablation on MM-GAG/A/G/T (

5\times 5

B=10

); each

C

column averages A/G/T.

Figure S1: Distance-stratified route gallery for DynCur-Geo. Each row contains two successes followed by two failures at the same initial distance. Every side cue rail shows the target photograph; failed panels also show the final-cell photograph. $S/F$ denotes success/failure, $T$ is path length for a success, and the pair on a failure is minimum/final distance.

Figure S2: Successful-route atlas for DynCur-Geo: four efficient successes, four detour successes, and four late-convergence successes. $C$ is initial distance, $S$ marks success, and $T$ is path length. Each panel pairs the route image with its target-cell photograph in a compact side cue rail.

Figure S3: Aggregate route-shape descriptors on all 735 shared diagnostic tasks per policy. DynCur-Geo denotes the proposed method, matching the main-paper terminology. The method has slightly higher path efficiency and progress-step ratio, and slightly lower revisit and regress ratios, than the two baselines.

Figure S4: Failure atlas for DynCur-Geo. Rows show four examples each of post-entry drift, revisit/ping-pong, approach-then-regress, non-convergence, and partial progress. Each compact side cue rail compares target and final-cell photographs, while a yellow cross marks the final cell in the full route view. $F$ marks failure; the numeric pair is minimum/final target distance.

Successful methods	$C=4$	$C=6$	$C=8$	Total
None	155	29	0	184
GeoExplorer only	57	15	1	73
GOMAA-Geo only	44	41	2	87
GOMAA-Geo + GeoExplorer	25	33	6	64
DynCur-Geo only	40	39	10	89
DynCur-Geo + GeoExplorer	29	40	8	77

Table S2: Success-combination counts on shared diagnostic tasks. Each row lists the methods that hit the target within

B=10

; omitted methods failed on the same task.

Figure S5: Four same-task cases in which DynCur-Geo succeeds and both baselines fail, shown as matched task columns and method rows. The columns cover $C=4,6,8$ while giving each route panel sufficient space. Start, target, and budget are identical within a column; the route panels omit target-photo insets so that the paths remain unobstructed. $S/T$ gives success and path length; $F/SG$ gives failure and terminal distance.

Figure S6: Four matched cases in which DynCur-Geo fails and at least one baseline succeeds, again shown as matched task columns and method rows without target-photo insets. The layout and annotations match Fig. S5 .

Figure S7: Matched-distance stepwise diagnostics. Each row compares a success (left) with a failure (right) at the same $C$ . Route panels include the exact target cue. Curves show remaining distance, the distance gate, and realized reward components. Successful routes contract the gate toward 0.405; failures show that regress actions reactivate the gate.

Figure S8: Step-level reward distribution on the seven recorded full-reward diagnostic routes used for the mechanism plots. The summary covers 58 transitions and groups them by whether the executed action reduces target distance, increases target distance, or reaches the terminal target cell. The plot decomposes each transition into external reward, gated intrinsic contribution, PBRS shaping, and total reward; it is a diagnostic decomposition rather than a full-route-bank estimate.

Figure S9: Controlled route comparison on one fixed $C=6$ task. The complete-reward policy reaches the target; the separately trained policies without the distance gate or without PBRS both finish at SG $=2$ . Only one training factor changes in each comparison.

Figure S10: Population dynamics over all 2,205 recorded routes. Bands and error bars are 95% percentile intervals from 2,000 deterministic bootstrap resamples clustered by aerial image. Top: mean remaining distance by initial distance. Bottom: cumulative exact-hit rate, progress probability conditioned on current remaining distance, and outcomes after first reaching distance one. Curves summarize the full trajectory bank.

Figure S11: Target/search protocol illustrations for aerial, MM-GAG, SwissView, and xBD inputs. The target cue is fixed before rollout, and the policy searches the corresponding observation stream.

Figure S12: Recorded xBD routes on four exactly matched tasks. Panels (a)–(d) show pre-only success, post-only success, success under both domains, and failure under both, respectively. Each panel pair fixes pair, start, pre-disaster target, $C$ , repeat, and budget; only the search image changes from pre-disaster (left route) to post-disaster (right route). Target-cue thumbnails are omitted so the route panels remain large and unobstructed.

Figure S13: Recorded cross-modal routes. Each row pairs an MM-GAG aerial target and route (left) with a SwissViewMonuments ground-photo target and aerial route (right). Cases cover long- and short-distance successes, near misses, and larger terminal errors. Swiss targets span castle, waterfront, civic-building, and dam/mountain scenes. The MM-GAG evaluator uses all 47 indexed samples; the Swiss rows use the controlled ground-to-aerial route set described in this supplement.

Recorded evaluator	Episodes	SR	SG	Succ. eff.	Revisit	Progress	Hit after $d=1$	Exit after $d=1$
xBD pre, aerial target	20,000	0.5973	1.3018	0.9304	0.1882	0.7585	0.7316	0.3465
xBD disaster, pre-disaster target	20,000	0.5984	1.2813	0.9322	0.1888	0.7599	0.7332	0.3475
MM-GAG, aerial target	1,175	0.6357	1.1319	0.9278	0.1732	0.7720	0.7477	0.3624
SwissMon ground, max80	1,380	0.5420	1.4457	0.8714	0.2176	0.7124	0.6997	0.4640

Table S3: Route-level descriptors recomputed from every saved trajectory in four evaluators. Succ. eff. is shortest distance divided by path length on successful episodes. The last two columns condition on first reaching distance one. The rows have different dataset scopes and are diagnostic rather than a controlled modality ablation.

Figure S14: Formal MASA enlarged-grid results. Panels (a)–(b) report per-distance SR for $8\times 8$ ( $B=24$ , $C=10,\ldots,14$ ) and $25\times 25$ ( $B=60$ , $C=12,16,\ldots,48$ ); panel (c) reports aggregate SR for $8\times 8$ , $10\times 10$ , and $25\times 25$ , where $10\times 10$ uses $B=32$ and $C=14,\ldots,18$ .

Training pool	Images	Avg. SR	Avg. SG
MASA+MM-GAG	184	0.6033	1.2543
MASA+SwissView100	237	0.5684	1.4191
MM-GAG+SwissView100	147	0.5497	1.7869
MASA+MM-GAG+SwissView100	284	0.5232	1.7952
SwissView100	100	0.5103	2.0490
MASA	137	0.4443	2.1380

Table S4: Training-pool selection under the unified evaluator. The final pool uses MASA+MM-GAG: 137 MASA images and 47 MM-GAG images. The comparison uses dynamically sampled navigation episodes from each image pool.

Scenario	Avg.	$C=4$	$C=5$	$C=6$	$C=7$	$C=8$
MASA/A	0.6200	0.2400	0.4000	0.6400	0.8800	0.9400
MM-GAG/A	0.6357	0.2979	0.3915	0.6894	0.8766	0.9234
xBD-pre/A	0.5973	0.3033	0.4065	0.6230	0.8115	0.8423
xBD-dis./A	0.5984	0.2923	0.4085	0.6290	0.8180	0.8443

Table S5: Distance-stratified DynCur-Geo results for the unified

5\times 5

B=12

C=4,\ldots,8

evaluator. Panel (a) reports SR and panel (b) reports SG. The same fixed task-bank protocol is used as in the unified main comparison.

Explore similar work

Sep 17, 2026

Towards Active Cross-View Object Geo-Localization

Cross-view object geo-localization (CVOGL) typically assumes a fixed query image, overlooking the ability of mobile agents to actively acquire more informative observations. To address this limitation, we introduce Active Cross-View Object Geo-Localization (ActiveGeo), where an agent sequentially selects new viewpoints and determines when to stop, aiming to improve localization with minimal observations. We further propose ActiveMoPT, an ActiveGeo framework with three-stage training. First, Multi-View Prompt-Preserving Adaptation enables the model to aggregate multiple query views while reusing the initial prompt. Second, Trajectory-Guided Policy Initialization uses supervised agent trajectories to learn viewpoint selection and initial stopping behavior. Third, Cost-Aware Policy Refinement employs GRPO with a gain-cost reward to jointly optimize localization accuracy and observation efficiency. We also construct ActiveGeo-858, a zero-shot test set containing 858 scenes and 1,716 target annotations. Experiments show that ActiveMoPT achieves state-of-the-art performance on MoP-UAV using only 1.45 query views on average, and substantially outperforms previous CVOGL approaches under zero-shot evaluation on ActiveGeo-858.

Shunyu Yao, Xiaohan Zhang, Zhuoran Yang +5

College of Information Science and Electronic Engineering, Zhejiang University, Hangzhou 310027, China · College of Computer Science, Beijing University of Technology, Beijing 100124, China · State Key Laboratory of Intelligent Green Vehicle and Mobility, Tsinghua University, Beijing 100084, China +2

Cross-View Geo-Localization Next-Best-View Planning

Ask PDF

Aug 27, 2026

UniGeo: A Multi-modal Large Language Model for Text-Guided Cross-View Geo-Localization

Text-guided drone geo-localization aims to identify a target region in a large-scale image gallery from a natural-language description. Existing methods mainly formulate this task as direct matching between an open-ended text query and candidate images. However, incomplete queries and highly similar candidates often make global cross-modal matching insufficient for reliable fine-grained localization. We propose UniGeo, a unified multimodal large language model (MLLM) for text-guided drone geo-localization. Built on a shared vision-language framework, UniGeo jointly supports geo-semantic understanding, cross-view semantic generation, and candidate-level verification. Specifically, it establishes stable correspondences among local scene elements, spatial relations, and language descriptions through geo-semantic learning, and further models semantic mappings between drone and satellite views through cross-view generation. Based on these capabilities, a plug-and-play verification module performs fine-grained discrimination among highly confusable candidates. We further introduce a multi-stage training strategy that progressively learns geo-semantic understanding, cross-view generation, and candidate verification, improving adaptation to text-guided geo-localization. Experiments demonstrate consistent improvements across multiple retrieval backbones. On GeoText-1652, UniGeo improves R@10 and mAP by 13.59 and 2.83 percentage points, respectively, validating its effectiveness for fine-grained text-guided drone geo-localization.

Jiahao Wen, Hang Yu, Zhedong Zheng

School of Computer Engineering and Science, Shanghai University, Shanghai 200444, China · Faculty of Science and Technology, and Institute of Collaborative Innovation, University of Macau, Macau 999078, China

Cross-View Geo-Localization Multimodal Large Language Models

Ask PDF

Aug 30, 2026

Perceive to Hypothesize, Verify to Ground: An Agentic Reasoning Framework for Open-World Geo-Localization

Open-world geo-localization requires models to reason over ambiguous visual cues through multi-step reasoning and external knowledge grounding. While recent large vision-language models exhibit strong multimodal reasoning capabilities, existing approaches still suffer from perceptual hallucination and context drift due to the lack of explicit evidence-grounded verification. In this work, we reformulate geo-localization as a human-like perceive-then-verify reasoning problem and propose GeoPAVE (Geo-localization Perception-and-Verification-Engine), a bi-level agentic framework that contains perception-based hypothesis generation via single-pass rollouts and verification-based evidence grounding for decision actions: support, refute, and refine. To support rigorous evaluation, we further introduce PAVED, a novel dataset derived from real-world user check-in data, equipped with comprehensive reasoning trajectories featuring multi-hop queries, multi-round tool invocations, and structured perception-verification traces. The dataset and code are available at https://github.com/Arandinglv/GeoPAVE.

Yutian Jiang, Ruijie Li, Sisuo Lyu +4

The Hong Kong University of Science and Technology (Guangzhou) · The Hong Kong University of Science and Technology

Multimodal Grounding Geospatial Reasoning Agentic Reasoning

Ask PDF