Robotic machine learning company Generalist has announced a new body AI system, GEN-1, that it says “reaches production-level success rates” for a “wide range of physical skills” that previously required human manual dexterity and muscle memory. Generalists are also touting the new model’s ability to “connect,” responding to disruption by improvising new movements.[ing] Gather ideas from different places to solve new problems. ”
GEN-1 is built on Generalist’s previous GEN-0 model, which the company touted in November as a proof of concept for the applicability of scaling laws in robot training, showing how increased pre-training data and computation time can improve post-training performance. But while large-scale language models have been able to effectively process trillions of words written all together on the internet as part of their training, robot models lack a similarly readily accessible source of high-quality data about how humans interact with objects.
To solve this problem, Generalist relied on “data hands,” a set of wearable pliers that capture minute movements and visual information as humans perform manual tasks. Generalist now claims to have collected over 500,000 hours of “petabytes of physical interaction data” to help train its physics models.
Shut up and take my money (out of your wallet) (and give it back).
The result is an autonomous system that’s precise enough to put money in your wallet and adaptable enough to fold laundry or sort car parts. According to Generalist, this model currently has a 99% success rate for repetitive but delicate mechanical tasks such as folding boxes, packing phones, and servicing robot vacuums, and is approximately three times faster than the previous GEN-0 model. The company says GEN-1 can achieve these marks with just an hour or so of pre-training adapted to the “robot data” applied to a particular robot embodiment.
recovery from mistakes
Until now, complex robotic systems have typically relied on carefully pre-programmed movements or been trained to focus only on a single task with little variation. What makes GEN-1 unique, according to Generalist, is that it can naturally respond to disruptions by improvising based on previous experience, even when a single model deviates “significantly from the training distribution.”
