AutoEvolve

Nebius Token Factory powers the cloud-side "task planner" by running Qwen 2.5 as a reasoning layer that takes in visual and natural language input — such as "pick up the cup" — and decomposes it into a series of fine-grained manipulation steps for robot arms, like approaching the object, opening the gripper, grasping, and lifting. A fine-tuned GR00T N1.6-3B consumes each step to produce continuous joint actions via its Vision Language Model (VLM), enabling the robot to perceive its environment, reason about what's happening, and execute meaningful physical actions across tasks like reach-and-grasp, object handoff, and basic manipulation.