an a single model enable a robot to do anything? One company is on a quest to find out.
Generalist AI, a startup aiming to build “physical AGI,” debuted GEN-1 last week, its latest attempt at creating a hardware-agnostic, general-purpose robotics model. The company said that GEN-1 has passed a new performance threshold, demonstrating the ability to master simple physical tasks with reliability, speed and improvisation.
According to Generalist, its model has a 99% success rate on tasks such as folding clothes, packaging items and folding boxes, whereas previous models only saw a 64% success rate. It also completes those tasks three times faster and requires only one hour of “robot data” per task, enabling “commercial viability across a broad range of applications.”
“While it cannot solve all tasks today, it is a significant step towards our mission of creating generalist intelligence for the physical world,” the company said in its announcement.
Generalist, which reached a $440 million valuation last year following a $140 million funding round, is one of several budding companies seeking to capture the momentum in physical AI. While many are focusing on specific form factors, like industrial arms, delivery bots or humanoids, Generalist’s model is designed not to be locked into any specific form factor.
At Nvidia GTC in March, Lucy Licht, Generalist’s partnerships lead, told The Deep View that the strength of the company’s model is that it’s “totally hardware agnostic.” By collecting an “incredibly diverse” set of pretraining data — one that the company has claimed is the world’s largest pretraining dataset for robotics — any use case, whether it be industrial, commercial, manufacturing or home, “exists inside of the model and just needs to be awoken by fine-tuning data.”
In a demo at the conference, a two-armed robot fitted with a Generalist model was able to delicately pick up a smartphone, place it in a small box and replace the lid.
“We see strong evidence and signs of life that the model works on a wide variety of embodiments,” said Licht.
Jamie Lee Solimano, the company’s applied AI lead, told me that the company’s work to-date has “ushered robotics into its pretraining era.”
“We've never before seen real scaling laws in robotics, because we've never had enough data,” said Solimano. “What we are seeing is an absolute revolution in robotics: To be able to show that pretraining works and completely reduce the amount of time that it takes to train up a new task or a new capability.”
Our Deeper View
Data is one of physical AI’s biggest roadblocks. Ken Goldberg, a professor at UC Berkeley, wrote a paper last year that detailed this problem, calling it the 100,000-year data gap: These models need far more data than is actually available. Generalist is on a mission to close that gap using pretraining, allowing robots to learn without the need for large simulation datasets. By vastly speeding up the process of training and deploying robots, those robots could then collect more data, get more training and become smarter. Hence, this creates the data flywheel that opened the same door in recent years for LLMs to grow as powerful as they are today.