Hybrid RL + Imitation Learning for Robotics with Ray

Learn how to implement hybrid reinforcement learning and imitation learning systems for robotics using Ray's distributed computing framework in this technical conference talk. Discover how to overcome the complex hardware and software challenges in robotics machine learning, where massive simulation parallelization requires RTX-class GPUs while policy training benefits from high VRAM and computational power. Explore a unified platform built on Ray that abstracts hardware and software diversity, allowing researchers to focus on scientific work rather than system administration. Understand the key features including unified orchestration through single Ray workflows for training full state RL models and multi-task IL policies, heterogeneous GPU scheduling with placement groups that automatically assign Isaac Lab simulators to RTX workers and gradient computation to A100/H100 trainers, and isolated deployment targets that package trained policies into lightweight Ray Serve microservices for robot deployment. Watch a live demonstration showing the launch of hybrid RL-IL runs with automatic provisioning of both Nvidia-RTX GPUs and A100/H100 nodes, observe Ray's adaptive cluster management as workloads shift between simulation, learning, and evaluation phases, and see real-time policy deployment to isolated robot runtimes. Gain practical design patterns for managing simulator-heavy workloads and large-scale network training within a reproducible Ray ecosystem, plus insights on meeting real-time robotics constraints while maintaining GPU efficiency.