Generalist AI's robots learn from videos, improvise with bananas and dustpans
A robot unzipped a purse, switched grippers when stuck, no task-specific training needed.
Deep Dive
At Generalist AI in Cambridge, MA, robot arms learned new chores from short instructional videos—no task-specific training required. In one demo, a robot improvised with a dustpan after its brush disappeared; another used a banana to sweep items. Founded by engineers who previously worked at Google DeepMind and Boston Dynamics
Key Points
- Robots learned tasks from short videos with zero task-specific training; one improvised with a dustpan when its brush was removed
- Founded by ex-DeepMind and Boston Dynamics engineers Pete Florence, Andrew Barry, and Andy Zeng; models built from scratch, not on open-source LLMs
- Training data collected via custom camera-equipped gripper gloves; hundreds of units headed to workers in Mexico to scale data collection
Why It Matters
Generalist AI's video-learned, improvising robots could finally make adaptable physical labor viable in commercial settings.