Generalist turns human demonstrations into robot policies for diverse cobot systems
Generalist, a robotics startup, leverages the Universal Manipulation Interface (UMI) research to train robots using human demonstrations captured with GoPro cameras and puppet-like end effectors. At Automate, the company showed how its models quickly generate policies for different cobot arms, including Universal Robots and Flexiv systems, with real-time error recovery. The approach aims to accelerate robot adoption across industrial applications.
Generalist's technology stems from a collaborative research effort involving Toyota Research Institute, Columbia University, and Stanford University. Their method captures human actions using GoPro cameras and puppet-style grippers, converting these demonstrations into training data. The startup has achieved a $2 billion valuation, operating within a sector that has collectively attracted over $4 billion in funding from competitors like Skild AI and Physical Intelligence.
At the Automate exhibition, the firm showcased its adaptability by programming Universal Robots arms to construct cardboard boxes while Flexiv systems repaired robotic vacuums. Castellanos emphasized that their hardware relies on a single-degree-of-freedom gripper, prioritizing simplicity and robustness over complexity to ensure reliable real-world performance.
By lowering the barrier to robot programming through human demonstration, this approach could accelerate automation adoption in small and mid-sized factories that lack specialized engineers. Workers may see their roles shift from manual execution to supervision and data collection. The emphasis on real-time error recovery could improve workplace safety and reduce downtime, though it may also raise concerns about labor displacement in repetitive tasks.