Bagel develops PARIS, a training architecture for robotics foundation models that enables different model components to learn independently before being composed into a unified system. In published tests with data and compute held equal, PARIS 1.0 improved results by 24% and PARIS 2.0 by 50%, though these results come from separate image- and video-generation tests rather than robotics validation.
Funding
Funding not disclosed
Founders
Product
Problem
Robotics models are typically trained as monolithic systems, requiring all components to be trained together with massive computational resources. This approach limits scalability, slows iteration, and makes it difficult to improve or swap individual capabilities without retraining the entire model.
Solution
Bagel provides PARIS, a training architecture that lets different parts of a robotics model learn independently and then be composed into a single unified model. This modular approach enables parallel development of individual components, reducing the computational burden and accelerating iteration cycles. In published tests with data and compute held equal, PARIS 1.0 improved results by 24% and PARIS 2.0 by 50%, demonstrating significant performance gains over monolithic training approaches. The architecture is designed as a general world action model, positioning it for broad application across robotics tasks that require perception, reasoning, and action generation.
Target Audience
Primary customers are robotics research labs and AI development teams building foundation models for robotic perception, planning, and action generation who need scalable training architectures.
Features
- Modular training architecture that decouples model components for independent learning and later composition
- PARIS 2.0 delivers a 50% improvement over baseline results in matched-resource tests
- General world action model design intended to support diverse robotics applications
- Published performance benchmarks from image- and video-generation tests demonstrating architecture efficacy