EBM-Nav: energy-based model for zero-shot visual navigation via implicit scene representation

Zero-shot visual navigation enables agents to generalize to unseen environments without task-specific training and is core to service robotics, warehouse logistics, and autonomous exploration. Existing approaches are observed to fail to achieve genuine zero-shot generalization: explicit-map methods rely on accurate reconstruction and are sensitive to sensor noise; supervised policies memorize training-scene patterns and degrade under illumination, furniture, and layout shifts; implicit representations offer continuous, differentiable encoding but remain confined to passive reconstruction and lack effective coupling with active decisions. To address this, EBM-Nav is proposed as a zero-shot navigation framework built on an energy-based model. Specifically, a periodic-activation implicit scene encoder is employed to map discrete RGB-D observations into a continuous, differentiable representation supporting arbitrary-resolution spatial queries. An energy function over the observation–action–goal triple is designed to reformulate navigation as gradient descent on an energy landscape. A frozen multimodal large language model is integrated to decompose long-horizon instructions into subgoals and generate candidate directions. A contrastive-divergence learning mechanism is developed with short-step Langevin dynamics to efficiently construct negative samples for training the energy function. Systematic experiments are conducted on the R2R-CE and RxR-CE continuous vision-and-language navigation benchmarks, and it is shown that EBM-Nav significantly outperforms state-of-the-art zero-shot methods in success rate, path efficiency, and navigation accuracy, while providing an interpretable energy-confidence mechanism for decisions. These results verify the energy-landscape zero-shot navigation paradigm, offer a general learning paradigm for embodied intelligence, and extend energy-based models from static perception to sequential decision-making, with methodological implications for service robots and autonomous exploration.