120 Minutes and a Laptop: Minimalist Image-goal Navigation via Unsupervised Exploration and Offline RL

Xiaoming Liu; Borong Zhang; Qingbiao Li; Steven Morad

cs.ROMar 27, 2026

120 Minutes and a Laptop: Minimalist Image-goal Navigation via Unsupervised Exploration and Offline RL

Authors: Xiaoming Liu, Borong Zhang, Qingbiao Li, Steven Morad

Organizations: Faculty of Science and Technology, University of Macau

Abstract

The prevailing paradigm for image-goal visual navigation often assumes access to large-scale datasets, substantial pretraining, and significant computational resources. In this work, we challenge this assumption. We show that we can collect a dataset, train an in-domain policy, and deploy it to the real world (1) in less than 120 minutes, (2) on a consumer laptop, (3) without any human intervention. Our method, MINav, formulates image-goal navigation as an offline goal-conditioned reinforcement learning problem, combining unsupervised data collection with hindsight goal relabeling and offline policy learning. Experiments in simulation and the real world show that MINav improves exploration efficiency, outperforms zero-shot navigation baselines in target environments, and scales favorably with dataset size. These results suggest that effective real-world robotic learning can be achieved with high computational efficiency, lowering the barrier to rapid policy prototyping and deployment.

Figures & tables

Fig. 1 : MINav enables rapid, two-hour deployment for real-world end-to-end ImageNav. (a) Autonomous data collection. (b) Offline policy learning. (c) Real-world policy deployment.

Fig. 2 : Real-world deployment of MINav in the three evaluation environments. The robot successfully reaches image-goals despite clutter, blur, reflections, and partial occlusion.

Fig. 3 : Overview of the MINav pipeline. Starting from autonomous exploration in the target environment, MINav collects diverse trajectories using the proposed pink uniform noise model, extracts visual representations with frozen DINOv3, constructs the goal space, and trains a navigation policy offline via hindsight goal relabeling and TD3+BC.

Fig. 4 : Generation process of the proposed pink uniform noise.

Fig. 5 : Valid goal selection using DINOv3-based spatial standard deviation.

Fig. 6 : Visualization of exploration trajectories generated by different noise strategies. The corresponding normalized state entropy $\eta(s)$ , action entropy $\eta(a)$ , and state-action joint entropy $\eta(s,a)$ are reported at the bottom of each subfigure, highlighting the superior coverage of the pink uniform strategy.

Fig. 7 : Simulation evaluation of downstream performance and offline model selection. (a) Success rates of offline policies trained on datasets collected with different exploration strategies. (b) Correlation between FQE scores and ground-truth success rates, showing that FQE is a reliable proxy for offline checkpoint selection.

Fig. 8 : Real-world evaluation environments. We consider three indoor environments of increasing difficulty: (a) Simple, (b) Standard, and (c) Complex.

Method	Visual backbones	Policy Params $\downarrow$	Training Cost (TFLOPs-hours) $\downarrow$	Simple			Standard			Complex
Method	Visual backbones	Policy Params $\downarrow$	Training Cost (TFLOPs-hours) $\downarrow$	SR( $\%$ ) $\uparrow$	Time( $s$ ) $\downarrow$	STL( $\%$ ) $\uparrow$	SR( $\%$ ) $\uparrow$	Time( $s$ ) $\downarrow$	STL( $\%$ ) $\uparrow$	SR( $\%$ ) $\uparrow$	Time( $s$ ) $\downarrow$	STL( $\%$ ) $\uparrow$
Random	N/A	N/A	N/A	$0\>\scriptstyle{\pm 0}$	$20\>\scriptstyle{\pm 0}$	$0\>\scriptstyle{\pm 0}$	$0\>\scriptstyle{\pm 0}$	$30\>\scriptstyle{\pm 0}$	$0\>\scriptstyle{\pm 0}$	$0\>\scriptstyle{\pm 0}$	$40\>\scriptstyle{\pm 0}$	$0\>\scriptstyle{\pm 0}$
NoMad	EfficientNet	$19\>M$	$>30000$	$18\>\scriptstyle{\pm 7}$	$18\>\scriptstyle{\pm 1}$	$14\>\scriptstyle{\pm 7}$	$18\>\scriptstyle{\pm 19}$	$25\>\scriptstyle{\pm 4}$	$16\>\scriptstyle{\pm 18}$	$18\>\scriptstyle{\pm 11}$	$37\>\scriptstyle{\pm 3}$	$16\>\scriptstyle{\pm 10}$
ViNT	EfficientNet	$30\>M$	$30000$	$8\>\scriptstyle{\pm 7}$	$6\>\scriptstyle{\pm 5}$	$19\>\scriptstyle{\pm 1}$	$13\>\scriptstyle{\pm 9}$	$28\>\scriptstyle{\pm 3}$	$9\>\scriptstyle{\pm 7}$	$8\>\scriptstyle{\pm 17}$	$38\>\scriptstyle{\pm 4}$	$7\>\scriptstyle{\pm 15}$
GNM	MobileNetV2	$9\>M$	$>5000$	$9\>\scriptstyle{\pm 12}$	$18\>\scriptstyle{\pm 1}$	$9\>\scriptstyle{\pm 11}$	$10\>\scriptstyle{\pm 10}$	$28\>\scriptstyle{\pm 2}$	$7\>\scriptstyle{\pm 10}$	$25\>\scriptstyle{\pm 13}$	$36\>\scriptstyle{\pm 3}$	$22\>\scriptstyle{\pm 11}$
Ours (Minimal, 1h)	DINOv3-ViT-S	$\textbf{2}\>M$	$150$	$73\>\scriptstyle{\pm 10}$	$12\>\scriptstyle{\pm 3}$	$60\>\scriptstyle{\pm 13}$	$43\>\scriptstyle{\pm 2}$	$24\>\scriptstyle{\pm 2}$	$29\>\scriptstyle{\pm 11}$	$25\>\scriptstyle{\pm 9}$	$33\>\scriptstyle{\pm 9}$	$21\>\scriptstyle{\pm 7}$

TABLE I : Real-world ImageNav performance across three environments.

Data Collection	Data Processing	Policy Training	Total Time
60 min	3 min (ViT-S)	22 min	85 min
60 min	22 min (ViT-L)	29 min	111 min
120 min	6 min (ViT-S)	22 min	148 min
120 min	44 min (ViT-L)	29 min	193 min

TABLE II : End-to-end runtime of MINav. The full pipeline is completed in under 120 minutes on a laptop.

Metric	SR ( $\%$ ) $\uparrow$	Time ( $s$ ) $\downarrow$	STL ( $\%$ ) $\uparrow$
White Uniform Noise	$15\>\scriptstyle{\pm 10}$	$26\>\scriptstyle{\pm 3}$	$14\>\scriptstyle{\pm 10}$
Pink Gaussian Noise	$30\>\scriptstyle{\pm 14}$	$24\>\scriptstyle{\pm 3}$	$27\>\scriptstyle{\pm 14}$
Pink Uniform Noise	$43\>\scriptstyle{\pm 2}$	$24\>\scriptstyle{\pm 2}$	$29\>\scriptstyle{\pm 11}$

TABLE III : Real-world ablation on exploration noise under a fixed 1-hour data collection budget.

Platform	Condition	SR( $\%$ ) $\uparrow$	Time( $s$ ) $\downarrow$	STL( $\%$ ) $\uparrow$
Quadruped (Go2)	Static	$100\>\scriptstyle{\pm 0}$	$13\>\scriptstyle{\pm 4}$	$92\>\scriptstyle{\pm 8}$
Quadruped (Go2)	Dynamic	$100\>\scriptstyle{\pm 0}$	$14\>\scriptstyle{\pm 2}$	$91\>\scriptstyle{\pm 3}$
Wheeled (Limo)	Static	$92\>\scriptstyle{\pm 18}$	$8\>\scriptstyle{\pm 3}$	$91\>\scriptstyle{\pm 18}$
Wheeled (Limo)	Dynamic	$88\>\scriptstyle{\pm 11}$	$14\>\scriptstyle{\pm 3}$	$83\>\scriptstyle{\pm 12}$

TABLE IV : Robustness of MINav across robot platforms and environmental conditions.

Robot Navigation

Robot navigation studies how robots perceive, localize, plan, and move safely toward goals in static or changing environments.

See all 554 papers

RoamFlow: Reinforcement-Aligned One-Step Action MeanFlow Policy for Image-Goal Navigation
Jun 29, 2026Zixuan Zhang, Yuqi Chen, Junjie Gao +4Visual Navigation Robot Navigation
Nanyang Technological University, Singapore
NavOL: Navigation Policy with Online Imitation Learning
May 12, 2026Xiaofei Wei, Chun Gu, Li ZhangVisual Navigation Robot Navigation
School of Data Science, Fudan University · 2Shanghai Innovation Institute.
FlashNav: Training Deployable Robot Navigation Policies in Seconds
Jun 14, 2026Shanze Wang, Yiwei Qian, Xinming Zhang +8Robot Policy Learning Reinforcement Learning
Eastern Institute of Technology, Ningbo · The Hong Kong Polytechnic University · National University of Singapore +2

Jun 29, 2026J/K move · Enter open · S save

RoamFlow: Reinforcement-Aligned One-Step Action MeanFlow Policy for Image-Goal Navigation

Zixuan Zhang, Yuqi Chen, Junjie Gao, Siyuan Song+3

Nanyang Technological University, Singapore

Visual Navigation Robot Navigation Image-Goal Navigation Goal-Conditioned RL

Open paper AskPDF arXiv

Image-goal navigation is a key challenge in embodied robotics, where an agent must reach a target specified solely by a goal image. While existing reinforcement learning approaches map perceptual observations directly to actions, they struggle to model long-horizon dependencies, often leading to suboptimal trajectories. To address this limitation, we propose RoamFlow, a generative navigation framework that leverages MeanFlow to predict the average velocity field for trajectory synthesis, enabling efficient few-step generation and reducing inference latency. We further adopt a two-stage training strategy that combines expert imitation for stable initialization with reinforcement learning for task-specific policy refinement. Extensive experiments in both Habitat simulation and real-world robotic platforms demonstrate that RoamFlow achieves efficient inference while maintaining strong navigation performance under real-time constraints.

Abstract

Robot Navigation

Explore similar work

RoamFlow: Reinforcement-Aligned One-Step Action MeanFlow Policy for Image-Goal Navigation

NavOL: Navigation Policy with Online Imitation Learning

FlashNav: Training Deployable Robot Navigation Policies in Seconds