Skip to content

Event

VEGA: Training Navigation VLAs from First-Person Videos

1 reports · 1 sources

Gershom Seneviratne and other authors propose VEGA, a method for training navigation vision-language-action (VLA) models from unlabeled first-person navigation videos, and simultaneously release the VEGA-Bench benchmark. The approach reconstructs local scene geometry from monocular video, samples text, image, and spatial waypoint goals and generates obstacle-avoiding trajectories, then uses these trajectories to train a flow-matching VLA navigation policy. The authors state that geometric information is used only during training, thereby distilling obstacle-avoidance planning capability into a purely visual policy. The authors state that the code and benchmark will be made public upon publication of the paper.

Report timeline

  1. VEGA: Training Navigation VLAs with In-the-Wild Egocentric Video and Geometric Trajectory Supervision

    VEGA proposes a method for training navigation Vision-Language-Action (VLA) models from unlabeled egocentric navigation videos. It reconstructs local scene geometry from monocular video, samples text, image, and spatial waypoint goals, and generates obstacle-avoidance trajectories, then uses these trajectories to train a flow-matching VLA navigation policy; geometry is used only during training.

    arXiv cs.RO ·

Related topics

Planning and decision makingRobot models and learningData and simulationResearch papers

Interest trend

Current
3
Peak
4Nov 4 01:00
Past 24 hours
–

The trend compares participants observed throughout the same period, so its range may be smaller than the current interest count. Move or tap to inspect each hour; use the arrow keys on a keyboard.