Bagel Labs is a physical AI research lab developing world-action models for robot control. We combine compact model architecture with efficient training to accelerate improvement across model generations.
PARIS adapts the topology of training to the distributed structure of the physical world. Specialized models learn independently, then compose into a unified model. This approach yields better results than monolithic training at matched data and compute.
WorldDiT is Bagel Labs’ compact model architecture for robot learning and control. It sits on the reported Pareto frontier for model size and task success. It is the foundation for one model designed to improve across tasks, environments, and robot types.
