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

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.

• The paper introduces the MINav pipeline, achieving rapid and efficient image-goal navigation using pink uniform noise and offline TD3+BC.
• The approach leverages a frozen DINOv3 encoder and a spatial standard deviation metric to effectively filter goal-relevant visual states.
• Empirical results show that MINav outperforms zero-shot baselines with high success rates in both simulated and real-world environments.

Minimalist, Rapid Image-Goal Navigation via Unsupervised Exploration and Offline RL

Motivation and Context

The study addresses image-goal visual navigation in real-world environments, emphasizing rapid, efficient deployment on specific robotic platforms with minimal computational resources and no human intervention. It critiques the dependence on large-scale pretraining and adaptation overheads in foundation models (e.g., GNM, ViNT, NoMaD), noting the practical bottleneck of domain specificity and real-time adaptation. The central research question is whether strong navigation performance can be achieved through efficient, purely in-domain learning, leveraging unsupervised autonomous data collection and offline RL.

MINav Pipeline

The proposed Minimalist Image-goal Navigation (MINav) pipeline comprises three fully automated stages:

1. Autonomous Dataset Collection: Utilizing a novel pink uniform noise model, the robot generates temporally correlated action sequences with uniform coverage across the action space, promoting maximal state-action exploration within tight real-world interaction budgets.
2. Offline Policy Training: Dataset coverage is filtered with a DINOv3-based spatial standard deviation (SSD) metric to construct a valid goal space. The RL state is defined by concatenated visual representations from four historical frames, approximating temporal dynamics under partial observability. Offline goal-conditioned RL is performed via TD3+BC with hindsight goal relabeling, mixing geometric trajectory sampling and global uniform goal selection.
3. Rapid Policy Deployment: The trained policy is deployed directly on the target robot in the environment, with checkpoint selection guided by fitted Q-evaluation (FQE), which has a demonstrated strong linear correlation (Spearman r = 0.91) with actual navigation success rates.

Technical Innovations

Pink Uniform Noise Exploration

• Coverage Optimization: Pink noise provides superior state exploration compared to white and Ornstein-Uhlenbeck (OU) noise, but standard Gaussian pink noise fails to adequately cover the action boundaries.
• Uniform Marginalization: Applying the probability integral transform to pink Gaussian noise enables uniform marginal distribution while preserving temporal correlation, leading to highest normalized entropy metrics (state, action, joint) in both simulation and real-world ablations.
• Impact: Under a fixed collection budget, pink uniform noise achieves substantially higher downstream navigation success rates than alternative noise strategies, demonstrating both broader and more balanced trajectory coverage.

Visual Representation and Goal Selection

• Frozen DINOv3 Encoder: Leveraging a frozen DINOv3 backbone allows for robust semantic encoding without the computational overhead of end-to-end training, facilitating efficient deployment.
• Spatial Standard Deviation: The SSD metric identifies visually informative goals, providing dense supervision and filtering out uninformative instances (e.g., frames captured near obstacles).

Offline Policy Learning

• TD3+BC: Policy optimization is regularized with behavioral cloning, minimizing out-of-distribution action estimation given exploratory datasets.
• Hindsight Goal Relabeling: Dense goal supervision is achieved through a mixture of geometric future and global uniform sampling, enabling more efficient training.
• Fitted Q-Evaluation: Enables offline checkpoint ranking without requiring real-world rollouts; this is vital for rapid deployment scenarios.

Empirical Results

Simulation Benchmarks

• Exploration Efficiency: The pink uniform noise strategy consistently achieves near-complete trajectory coverage and highest entropy metrics across all budgets.
• Scaling Properties: Success rates increase monotonically with data budget, e.g., 1 hour pink uniform dataset comparable to 3 hours pink Gaussian.

Real-World Deployment

• Minimalist Pipeline Performance: The full MINav pipeline, from autonomous data collection to deployment, completes within 120 minutes on a consumer laptop, confirming operational practicality.
• Benchmark Comparison: MINav, trained on 1–2 hours of in-domain data, decisively outperforms zero-shot baselines (GNM, ViNT, NoMaD), which utilize 70+ hours of pretraining, with up to 100% success rate in static and dynamic environments.
• Data Scaling: Increasing training budget from 1 to 2 hours yields consistent performance improvements across all metrics and environments.
• Visual Encoder Ablation: Compact ViT-S DINOv3 backbones perform on par with larger variants except in highly complex environments, validating efficient resource usage.
• Noise Ablation: Pink uniform noise yields the highest success and STL metrics under fixed collection budgets, attributable to superior trajectory diversity.
• Cross-Platform Robustness: MINav adapts seamlessly to both quadruped and wheeled robots without pipeline modification, maintaining high SR (>88%) and robust performance under dynamic human interference.

Implications and Future Directions

Practically, the results indicate that offline, in-domain RL pipelines (such as MINav) with efficient action-space exploration and robust visual encoding can circumvent the prohibitive cost and low deployment efficiency of foundation model adaptation for navigation tasks in unmapped environments. The approach scales favorably with data, is robust to embodiment changes and environmental perturbations, and requires minimal computational resources. Theoretically, the findings encourage further exploration of colored noise exploration in offline RL and the utility of frozen transformer-based encoders for robotic tasks. Future directions encompass stronger generalization under limited on-policy data, extension to vision-language navigation, and adaptation to multi-agent or multi-objective settings.

Conclusion

MINav demonstrates that a minimalist, fully automated framework—based on unsupervised exploration, frozen visual representations, and offline goal-conditioned RL—can rapidly deliver robust end-to-end navigation policies. Leveraging high-quality, temporally correlated datasets and efficient offline learning, MINav attains superior deployment efficiency and performance in real-world scenarios, outclassing zero-shot navigation baselines, scaling with data, and showing robustness across platforms and dynamic conditions (2603.26441). This approach substantively lowers the barrier for practical, rapid prototyping and deployment of navigation policies in robotics.