Co-Ops: Collaborative Open Source and Privacy-Preserving Training for Learning to Drive
Note: This project is a continuation of OSMOSIS: Open-Source Multi-Organizational Collaborative Training for Societal-Scale AI Systems.
Current development of autonomous vehicles, a socially transformative technology, has been slow, costly, and inefficient. Deployment has been limited to restricted operational domains, e.g., a handful of cities and routes, where systems often fail to scale and generalize to new settings such as a new vehicle, city, weather, or geographical location. How can we enable urgently needed scalability in the development of autonomous driving AI models?
To address this critical need, our project introduces Co-Ops, a novel framework for collaborative development of training AI models at scale. Co-Ops will work towards enabling generalized, seamless and accessible collaboration among individuals, institutions and companies through two main innovations. First, we will develop a standardized and privacy-preserving platform for flexible and incentivized participation in distributed AI model training. Second, we will design principled AI models and training techniques that can effectively learn from unconstrained and heterogenous data, including different geographical locations, vehicles, rules of the road, social norms, and traffic regulations. With these two innovations, Co-Ops will enable high-capacity and open-source models that can easily scale across various locations and conditions, from Boston’s narrow roads to Singapore left-hand drivers.
Figure caption: High-level description of a holistic platform and research challenges that will be addressed in the proposed mass collaborative open-source framework. As scalable training over heterogeneous settings is not feasible with current collaborative AI training frameworks, we will introduce and explore novel client and server-based modules for joint privacy and driving model safety optimization (i.e., model selection, validation, capacity).