Pete Florence, co-founder and CEO of Generalist AI, discusses scaling laws and emerging capabilities in robotics foundation models at Fellows Forum.
Florence describes Gen 0 as a test of whether more general physical data could improve many tasks, rather than only the task represented by that data. He says the models improved with more compute and data, including on 16 downstream tasks outside pretraining. The dataset contained 270,000 hours, and real-world success on one challenging task reached 57%, with large error bars.
Gen 1 used roughly an hour of task-specific robot data to reach success rates above 99% across varied tasks, including folding shirts and automotive kitting. Florence highlights faster execution and an example where a robot used its second hand to reposition a washer, drawing on behavior learned during pretraining rather than that task’s fine-tuning.
Gen 1.5 can take a robot or human-glove demonstration as an in-context prompt and attempt tasks such as sweeping blocks, opening a jar, or pouring bolts. Florence also shows chained tasks, live imitation, and adaptation through as few as one to ten gradient steps. In one example, the robot misses a cup on its first attempt and then corrects its action.
Florence shows the model using a banana as a brush and a dustpan to move a block into a bowl, then describes other unplanned behaviors such as clearing a Lego brick from a finger. In the Q&A, he says the models work across many hand designs and can generalize between arm pairs, while hands remain the hardest embodiment variation. He views touch as useful for some tasks but not the main barrier to general robot intelligence, and says stronger pretraining reduces the need for reinforcement learning that starts from scratch.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Pete Florence of Generalist discusses the path toward robots that learn physical tasks from demonstrations and describes the company’s Gen 1.5 model at Actuate 26; Foxglove is the channel.
