Paper review · October 10, 2026
SuperNav: An Agentic Navigation System That Works for Any Task in Any Scene
SuperNav: An Agentic Navigation System for Any Task in Any Scene
By wrapping a pretrained MLLM, without fine-tuning, in a navigation-specific agent harness (Skills, Tools and context management), the system achieved higher success rates than the 4 baselines it was compared against across several navigation tasks.
Problem
Service robots must handle varied requests in unfamiliar environments. Approaches that fine-tune an MLLM to predict actions directly are bound to the scope of their training data and generalize poorly to new tasks and scenes, while modular zero-shot systems require modifying fixed workflows for each task.
Method
The MLLM handles only request interpretation, scene understanding and decision-making, leaving movement execution to tools. An agent loop repeats observation, movement, goal-progress management and termination Tool calls, and processed images are removed from later requests, leaving only their paths, to reduce context. The MLLM specifies a destination as a single point on 4-direction observation images (PointNav), which a geometry-based executor using depth and pose, or a learned executor, carries out. Navigation Skills for exploration, recovery and completion judgment are Markdown documents read when needed. Experiments used GPT-5.6 Terra.
Results
- On 150 single-object navigation episodes, it reached an SR of 78.00%, higher than the strongest baseline, UniNaVid (34.00%) (Geo-based Executor).
- On 150 ordered multi-object navigation episodes, it reached an SR of 34.00%, higher than the best baseline, OmniNav Action Former (4.00%).
- On 200 demand-driven navigation episodes based on Demand-Bench, it reached an SR of 59.50%, higher than OmniNav Action Former (37.50%).
- On 120 episodes of HM3D-OVON val-unseen, without OVON-specific training, it recorded SR 73.33% and SPL 0.4105 at the 1 m threshold, and SR 68.33% and SPL 0.3837 at the 0.25 m threshold.
- The authors report that it was also deployed on a real quadruped robot and shown to work.
Limitations
The instance benchmarks were built by the authors themselves, and the demand-driven evaluation used a modified protocol. The baselines are mostly MLLM fine-tuning approaches, so comparison conditions differ. HM3D-OVON was evaluated on a 120-episode subset. Multi-object SR is still low at 34%, and the authors state that long-sequence demand-driven tasks remain difficult. Performance depends on a commercial MLLM (GPT-5.6 Terra).
Why it matters
It is a case of carrying harness design (tools, skills and context management) that proved effective in general-purpose agents over to robot navigation. It shows a direction that uses a general-purpose model's capabilities as they are, without separate training on navigation data.
Key terms
- Agent Harness
- A framework that provides an execution loop, tools and context management so a model can call tools, receive results and decide again.
- Visual-point interface
- An approach in which the MLLM picks a destination as a single coordinate point on an observation image, and a movement tool plans the actual path and moves there.
- SPL(Success weighted by Path Length)
- A metric that multiplies success by the efficiency of the actual path relative to the shortest path. Even a success scores low if it takes a detour.
- Navigation Skill
- Markdown instructions describing tips for exploration, recovery and completion checks. The model loads them into context and consults them when needed.
Original
Authors: Jinkai Zhang, Jingyi Xu, Yuanhong Yu, Jiarui Guo, Ruizhen Hu, Hujun Bao, Xiaowei Zhou, Sida Peng · Posted 2026-10-08 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.