DA-Nav gives robots recovery-aware long-horizon navigation and 98.15% correction success

新智元 · wechat · 2026-07-22

DA-Nav teaches robots not just to follow routes, but to recover after drifting off them

StarSource Intelligence introduces DA-Nav, a direction-aware visual-language navigation framework for city-scale long-horizon navigation. The key idea is to unify route instructions, first-person spatial understanding, and drift recovery into one decision chain.

Instead of directly predicting continuous 3D trajectories, the system first identifies candidate spatial regions in the robot’s camera view and then converts those region-level decisions into executable motion. The navigation loop is structured as:

To address the fact that real robots often drift off the planned route, the team built ReDA, a recovery-aware dataset with controlled perturbations. It contains about 286k temporal samples, including about 128k recovery samples. The model learns three things explicitly:

Reported results include a 98.15% correction success rate, 59.00% navigation success rate, and 77.82% route completion rate in long-horizon closed-loop evaluation. The same high-level policy was also deployed on quadruped and humanoid robots and validated over 1,200+ meters of real-world navigation without fine-tuning on real-world data.

The broader contribution is shifting long-range navigation from a path-generation problem to a continuous execution-and-recovery problem.

Original post →

More from Embodied

Embodied channel →