🦾SuperNav Drives a Robot Dog With an Unmodified MLLM
TL;DR
SuperNav keeps a pretrained multimodal LLM frozen and wraps it in an agent harness with navigation skills and tools. The model decides, tools move.
SuperNav keeps a pretrained multimodal LLM frozen and wraps it in an agent harness with navigation skills and tools. The model decides, tools move. It beat four baselines and ran on a real quadruped robot.
Key Points
No navigation-specific fine-tuning; the MLLM interprets requests and scenes
Harness supplies Navigation Skills, physical-interaction Tools and task-progress management
A visual-point interface lets the model pick destinations in images and revise from execution feedback
Beats four baselines on instance-level, multi-object and demand-driven tasks
Category-level tests on HM3D; deployed on a real quadruped
Why It Matters
The harness-over-fine-tuning pattern from coding agents is reaching robotics. It lets teams ride frontier model upgrades without retraining.
Quick Facts
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,564 builders reading daily.