Researchers Flag Hardware and Verification Limits in Robotics
Sara Hooker and Animesh Garg discuss hardware, data, and verification hurdles slowing robotics relative to general AI.
Sara Hooker posted that robotics advances at a different pace than general AI because it cannot access the same data catalysts. She noted optimism from cheaper synthetic data narrowing gaps and progress on non-verifiable rewards benefiting multiple fields. Animesh Garg replied that stable hardware interfaces and agreed pre-training standards remain major obstacles, with real-world verification often hidden behind IP. Sara Hooker and Animesh Garg made these points in replies on X.
It is really interesting to understand this dynamic, because it also speaks to the core problems our current AI doesn't solve. Robotics is progressing at a different pace because it can't tap cleanly into the same catalysts + data that have moved general AI at hyperspeed.
Combined views
Researchers Flag Hardware and Verification Limits in Robotics
Sara Hooker and Animesh Garg discuss hardware, data, and verification hurdles slowing robotics relative to general AI.
Sara Hooker posted that robotics advances at a different pace than general AI because it cannot access the same data catalysts. She noted optimism from cheaper synthetic data narrowing gaps and progress on non-verifiable rewards benefiting multiple fields. Animesh Garg replied that stable hardware interfaces and agreed pre-training standards remain major obstacles, with real-world verification often hidden behind IP. Sara Hooker and Animesh Garg made these points in replies on X.
It is really interesting to understand this dynamic, because it also speaks to the core problems our current AI doesn't solve. Robotics is progressing at a different pace because it can't tap cleanly into the same catalysts + data that have moved general AI at hyperspeed.
Reactions from ranked influencers
2 postsWhile not proceeding at the same pace, some reasons to be optimistic pace of progress for robotics is accelerating. - we are able to build synthetic data more cheaply and easily now. So the data gap may not be as severe as in previous ML eras. - solving non-verifiable problems…
@sarahookr Stable standard hardware interface and lack of agreement on the right pre-training has been a big challenge Verification exists in the real world but is often behind IP
