Last week, I visited Generalist AI’s offices in Cambridge, Massachusetts, where I observed robot arms performing simple chores like stacking cups and putting blocks into bowls. The robots mastered these tasks after watching a short instructional video, with no prior training for each specific activity. One robot, instructed to sweep a block into a bowl using a dustpan and brush, improvised by using the dustpan like a brush when the brush was removed from the scene. This ability to adapt and improvise was particularly striking.
The company’s approach appears to focus on teaching robots about the physics of the world, inspired by how humans intuitively understand physical interactions from an early age. This may help robots transfer knowledge between different tasks. For example, a robot once used a banana to sweep up items when presented with the object. While this may seem trivial, it highlights the gap in physical intelligence among machines. Researchers have often been surprised by the creative decisions robots make during tasks, which suggests a growing ability to experiment and improvise.
Generalist AI’s cofounders, Pete Florence and Andrew Barry, have backgrounds at Google DeepMind and Boston Dynamics, working on advanced robotic models. The company is building a general robotic model trained by humans, using special gloves with cameras that people use to perform chores. These gloves are destined for workers in Mexico and other regions. While the company is cautious about its training methods, it claims to have gathered a large amount of high-quality training data, building its AI models from scratch rather than relying on open-source language models.
Source: wired
