Autonomous Systems offer hope towards moving away from mechanized, unsafe, manual, often inefficient practices. Last decade has seen several small, but important, steps towards making this dream into reality. These advancements have helped us to achieve limited autonomy in several places, such as, driving, factory floors, surgeries, wearables and home assistants, etc. Nevertheless, autonomous systems are required to operate in a wide range of environment with uncertainties (viz., sensor errors, timing errors, dynamic nature of the environment, etc.). Such environmental uncertainties, even when present in small amount, can have drastic impact on the safety of the system—thus hampering the goal of achieving higher degree of autonomy, especially in safety critical domains. To this end, I shall discuss formal techniques that are able to verify and design autonomous systems for safety, even under the presence of such uncertainties, allowing for their trustworthy deployment in the real world.

Specifically, I shall discuss monitoring techniques for autonomous systems from available (noisy) logs, and safety-verification techniques of autonomous system controllers under timing uncertainties. Secondly, using heterogeneous learning-based cloud computing models that can balance uncertainty in output and computation cost, I will present techniques for designing safe and performance-optimal autonomous systems. I will conclude this talk by outlining a number of future research directions on how we can build a network/environment where several autonomous systems coordinate among themselves and achieve their goals in a safe and performance optimal manner even in presence of uncertainties.