How WPP accelerates humanoid robot training 10x with G4 VMs

WPP's Creative AI team shares how they trained humanoid robots to perform complex dance sequences using a reinforcement learning workflow on Google Cloud G4 VMs powered by NVIDIA RTX PRO 6000 Blackwell GPUs. By combining OptiTrack motion capture, OpenUSD digital twins, MuJoCo physics simulation, and NVIDIA Isaac Sim, they reduced training cycles from 24 hours to under one hour — a 10x+ speedup. The post covers the sim-to-real gap challenge, how billions of simulations produce ONNX policies deployed to physical robots, and points to open-source Unitree Robotics RL code on GitHub for others to replicate the workflow.