Robot Deployment Data Limits Progress Toward General-Purpose Policies
Reactions from ranked influencers
3 posts
How much is robot deployment data worth? Deploying robots would seem to be a good way of working towards a general-purpose policy -- but I think in most cases, it is not. Why? Because to get to the high success rates you need for deployments, you constrain the problem too much; lacking novelty, you learn a bunch of different islands of disconnected data which fundamentally don't help work towards generality, and won't help onboard new tasks. A short blog post:
First off, partly inspired by this post from @KyleVedder of Physical Intelligence from a while back:
opinion piece: cargo cults, data flywheels, and novelty pumps tl;dr: “deployment data flywheels” miss that scale should provide novelty, and deploying robots on a known task wrings most novelty out of its generated data your ops team’s novelty pump is the only way out 🔗👇
@KyleVedder How Much Is Robot Deployment Data Worth? On robot deployment flywheels and why I am skeptical of the data they produce Full blog post: https://open.substack.com/pub/itcanthink/p/how-much-is-robot-deployment-data?r=3uj5js&utm_campaign=post&utm_medium=web