Jaival Patel

hot take, aerospace can grow quicker if we trusted reinforcement learning more in the domain. there are many ways we can impose safety-critical RL: - control barrier functions: keeping the policy inside a mathematically safe region (surrogate model?). RL agent can optimize performance, but a barrier layer can block actions that violate constraints. - run time safety filters - shielded RL: a policy that proposes actions, but a classical controller or rule-based shield that overrides unsafe actions. - hybrid RL + classical control (what i worked on last month) - uncertainty-aware policies: using ensembles, Bayesian models, or confidence estimates so the system knows when it is uncertain. (im looking into this right now for commercial robotic systems and it is giving me good results so far) in the end, RL doesn't have to replace classical GNC. it simply just needs to sit inside the constraints - it's RL inside safety-critical control architecture.