Generalist GEN-1.5 makes robots one-shot learners

GEN-1.5 makes robotics feel less like programming and more like teaching, but its short-horizon tasks and brittle in-context performance keep this at breakthrough demo stage.

• –Physical prompts combine sensor data and action trajectories, letting robots infer tasks from demonstrations rather than language alone
• –Two demonstrations can be composed into longer behaviors with new repositioning, regrasping, and recovery motions
• –Simulation and human demonstrations can transfer directly to real robots, potentially reducing costly robot-data collection
• –Few-step adaptation cuts task-specific training from thousands of updates to as little as one gradient step
• –The reported 59% one-shot success rate is promising, but production reliability remains unproven