Large-scale robot learning, such as using large behavior models (LBMs), has increasingly become the norm in robot learning literature. Nevertheless, many questions remain about how to deploy these embodied learning-based systems in the real world. First, we need a rigorous and efficient means to evaluate candidate robot learning models with statistical guarantees. Second, to deploy these models in human environments, they should be equipped with reliable failure detection systems despite the challenge posed by countless and often unforeseeable failure types and conditions during deployment. Lastly, these models should have the capacity to explore and adapt to new environments, preferences, and tasks. In this talk, I will overview our work along these research directions as part of the Trustworthy Learning under Uncertainty (TLU) team at TRI, focusing on rigorous and fast sequential evaluation, detecting failures on the fly, and adapting large behavior models to new tasks during deployment. These research lines aim to spur algorithmic development for robust and reliable deployment of embodied foundation models in human environments. To learn more about our work, visit: https://www.tri.global/trustworthy-learning-under-uncertainty-tlu.

Learning-based controllers now achieve impressive performance on hard robotics and autonomy tasks, but two coupled gaps limit their deployment: we cannot guarantee that a learned policy is safe, and we cannot efficiently find the rare scenarios in which it fails. This talk presents our recent development that closes the loop between synthesis and falsification. On the synthesis side, I will describe how neural certificates can be formulated as value functions in reinforcement learning to produce control policies with safety and stability guarantees, turning safety from a soft penalty into a structural property of the policy. On the verification side, I will present an adaptive test-case generation approach that uses Bayesian active learning to discover scenarios that are both likely to reveal failures and useful for coverage. This method can predict failures from limited hardware demonstrations and support system-level ethical testing through scalable, evolving experimental design.

Skydio’s autonomy enables flying robots to navigate complex environments, track subjects, and execute missions through high-level operator commands. This talk explores the full autonomy stack powering Skydio’s flying robots, spanning the cloud and the edge. The recently unveiled F10 Lightrunner, an aircraft with a distinctive asymmetric design, extends autonomy to long-range fixed-wing flight and robotic takeoff and landing from MegaDock. These capabilities enable a single operator to orchestrate multiple drones remotely through a web browser, bringing autonomous robotics to public safety and critical infrastructure.