Parallel Differentiable Reachability for Learning and Planning with Certified Neural Dynamics and Controllers
Researchers have developed a new parallelizable, differentiable reachability framework designed for continuous- and discrete-time systems. This framework integrates Taylor-model flowpipe construction with linear bound propagation, enabling GPU-batched computation and automatic differentiation. The system supports both analytical and neural network-based dynamics and controllers, offering a way to provide formal guarantees under uncertainty for closed-loop neural systems in robotics. AI
IMPACT Enables formal guarantees for neural network-based robotics systems, potentially improving safety and reliability in complex tasks.