- Keywords: Planning, Model learning
- TL;DR: Real-Time , Zero-Order Optimization with Learned Models
- Abstract: For the real robot challenge, we combine a fast zero-order optimization method (iCEM) with model learning. Our iCEM method generates strong expert trajectories using a learned forward dynamics model, which is iteratively trained on newly collected data. We use parallel forward model ensembles and use their uncertainty as a cost penalty in the optimization. Through the tight coupling between all components, our algorithm runs in real-time, rendering it suitable for real robotic applications.