Assurance for Learning Enabled Systems
This project will extend previous work in automatic synthesis methods for planning and model-predictive control to address (1) real-time synthesis through efficient and incremental constraint-solving, and (2) risk-awareness by explicitly modeling uncertainty in perception and dynamics modeling. ALES includes a risk-aware planning and real-time synthesis engine to generate plans, protocols and control that are correct-by-construction. We plan to use a special class of stochastic predicates called chance constraints to express confidence in the learned component or its individual outputs. We specify the semantics of the temporal evolution of this logical model based on an underlying learning algorithm.
SRI and Caltech shall develop algorithms for correct-by-construction synthesis from high- level contracts, and planning in presence of uncertainty. Caltech's primary objectives are to support the following milestones:
- Year 1: Develop algorithms for correct-by-construction synthesis with probabilistic notion of safety
- Year 2: Extend the approaches to incorporate risk measures such as Conditional Value-at-Risk.
- Reactive motion planning with probabilistic safety guarantees. Yuxiao Chen, Ugo Rosolia, Chuchu Fan, Aaron D. Ames, Richard M. Murray. Submitted, Conference on Robotic Learning (CoRL).
- Counter-example Guided Learning of Bounds on Environment Behavior. Yuxiao Chen, Sumanth Dathathri, Tung Phan-Minh, Richard M. Murray. 2019 Conference on Robot Learning (CoRL).
The project or effort depicted was or is sponsored by the Defense Advanced Research Projects Agency (Agreement FA8750-19-C-0089). The content of the information does not necessarily reflect the position or the policy of the Government, and no official endorsement should be inferred.