AWS and NVIDIA enable robot reinforcement learning training on SageMaker AI using Isaac Lab simulation framework

AWS announced integration of NVIDIA Isaac Lab, an open-source robot learning framework, with Amazon SageMaker AI for training robot policies. The solution supports training across two compute options: SageMaker HyperPod (for long, distributed training runs) and SageMaker Training Jobs (for iterative experiments). Example task: training a Unitree H1 humanoid robot to track velocity commands while walking across rough terrain using 19 coordinated joints.

SageMaker HyperPod adds managed cluster resiliency with automatic node health checks, fault detection, and auto-resume from the last checkpoint with no manual intervention. SageMaker Training Jobs provide ephemeral, on-demand compute that provisions instances, runs training, uploads artifacts, and terminates—eliminating idle compute costs between runs. Both use the same Docker training image built from NVIDIA Isaac Sim 5.1.0 with Isaac Lab v2.3.2.

The approach lets robotics teams compress what would take months of real-world training into hours of GPU-accelerated simulation by running thousands of robot instances simultaneously on one or multiple GPUs. Training metrics stream to Amazon SageMaker managed MLflow for persistent, searchable experiment tracking across both backends when configured.