Generalist AI has released GEN-1.5, a system that allows robots to learn new actions from a single video demonstration without prior training. The model accepts a clip lasting between three and twelve seconds as a physical prompt within its context window. After viewing this example, the robot attempts to replicate the movement immediately. Testing across ten simple tasks, such as opening a jar or retrieving cash from a wallet, showed an average success rate of 59 per cent. When the company applied ten training steps using five minutes of data, that figure rose to 83 per cent. The technology can also chain multiple prompts together and imitate human hand movements partly. Generalist states these capabilities emerged naturally during eight months of pretraining on interaction data rather than through explicit instruction. While other teams have demonstrated similar in-context learning for limited task types, this model claims broader applicability. However, the demonstrated tasks remain simple and short, and all results originate from the company itself without independent verification.
The significance lies in the potential to reduce the data required for robot programming, though current limitations prevent broad adoption. The approach relies heavily on unverified internal testing and simple scenarios.
* Average success rate of 59 per cent without training data
* Success rate increased to 83 per cent with ten training steps
* All results remain unverified by external researchers
