Event
Research on Real-Time Navigation Safety Representation Based on Reachability Released
1 reports · 1 sources
Bojan Derajić released a thesis proposing the use of neural implicit representations to offline-learn the value function of Hamilton-Jacobi (HJ) reachability analysis, and to provide it to optimization-based controllers during real-time navigation, in order to address the insufficiency of geometric collision checking alone in unknown dynamic environments. For control-affine dynamics and polytopic inputs, the method exploits the property that the HJ Hamiltonian attains its maximum at vertices, generating training labels through a parallelizable vertex-input tree search; combined with an HJ variational inequality loss, it can obtain barrier candidates for systems with up to 13 states. Simulations and hardware experiments on ground and aerial robots show that its collision-avoidance performance is superior to geometric and hand-designed baseline methods, especially in short-horizon and moving-obstacle scenarios.
Report timeline
- Reachability-Learning-Based Safety Representations for Real-Time Robot Navigation in Unknown Dynamic Environments
This dissertation uses offline-trained neural implicit representations to provide the value function from Hamilton-Jacobi (HJ) reachability analysis in real time to optimization-based controllers, addressing the inadequacy of geometric collision checking in unknown dynamic environments.
arXiv cs.RO ·