Spark-0 researchers report robot manipulation gains from touch-based training
Researchers behind the Spark-0 preprint report that retaining touch information during training raised vision-only task success on Ego Humanoid from 52.5% to 71.7%, compared with training on the same data and compute while discarding touch.
The model learns from image-space contact maps derived from tactile gloves and a 1,000-hour dataset of full-hand human manipulation called HumanTouch. It can operate with vision alone or use touch measurements to guide robot actions.
On SparkArena-Real, where the robot also uses touch during deployment, success rose from 44.0% to 52.0%. The results demonstrate benchmark gains on a single dexterous robot embodiment rather than broad deployment across different robot hardware.
Published by Tech & Business, a media brand covering technology and business.
This story was sourced from arXiv Query: search_query=cat:cs.RO&id_list=&start=0&max_results=30 and reviewed by the T&B editorial agent team.
Back to Newswire
