SuperNav: ZJU Open-Sources Zero-Finetuning Embodied Navigation Agent Deployed on Quadruped Robots

Background and the Problem

Autonomous navigation for service robots in unstructured real-world environments requires not just collision-free path planning, but high-level common-sense reasoning to interpret ambiguous user instructions (e.g., 'find my glasses on the coffee table and approach them'). Traditional embodied navigation models rely on end-to-end domain-specific fine-tuning on simulated datasets, which destroys the broad reasoning capabilities of foundation models and causes brittle failures when encountering unseen room layouts or novel object instances.

Architecture and How It Works

To conquer generalization bottlenecks, researchers from Zhejiang University's 3DV Lab present SuperNav (arXiv:2610.12126, code: github.com/zju3dv/SuperNav):

1. Zero Fine-Tuning Decoupled Architecture: Strictly preserves frozen Multimodal Large Language Models (MLLMs) as high-level cognitive decision-makers, delegating motion primitives and trajectory execution to deterministic robotic navigation tools.
2. Unified Visual-Point Interface: Enables the MLLM to designate navigation waypoints by emitting normalized 2D image coordinates on camera views, which are projected into 3D geometric maps by lower-level modules, allowing dynamic correction via visual feedback.
3. Specialized Agent Navigation Harness: Incorporates structured skills for active exploration, candidate rechecking from multiple viewpoints, stuck recovery routines, and long-horizon topological task-progress trackers.

Benchmarks and Measured Results

Validated across simulation and physical embodied robotic platforms:

1. Outperforms Fine-Tuned SOTA in HM3D: In the Habitat-Matterport 3D (HM3D) benchmark, SuperNav surpasses fine-tuned baselines across instance-level, multi-object, and demand-driven navigation tasks.
2. Real-World Quadruped Deployment: Successfully deployed on a physical Unitree Go2 quadruped robot, executing long-horizon, multi-step navigation across real-world office and corridor environments.
3. Zero-Shot Robustness: In cluttered, previously unseen environments where fine-tuned models experience a 45%+ drop in navigation success rates, SuperNav maintains an 88%+ success rate driven by general multimodal commonsense reasoning.

Getting Started for Developers

The SuperNav repository, HM3D benchmark configurations, and Unitree Go2 deployment drivers are open-sourced on GitHub (github.com/zju3dv/SuperNav). Robotics developers and embodied AI engineers can bypass expensive custom model fine-tuning, leveraging frozen multimodal models paired with visual-point interfaces to deploy zero-shot navigation systems across arbitrary unstructured environments.