#Tag (1)

VIRAL: Visual Sim-to-Real at Scale for Humanoid Loco-Manipulation

VIRAL from NVIDIA and CMU: 200 simulated teleoperation demos serve as reference state initialization for a privileged RL teacher (16 GPUs), then a DAgger+BC mixture distills an RGB-only student across 64 GPUs with tiled rendering, deployed zero-shot on a Unitree G1 with no real-world fine-tuning. It succeeds in 54 of 59 consecutive real trials at a 20.2 s cycle time, faster than an expert teleoperator with 1000+ hours of experience (21.4 s). Key finding: compute scale is a requirement, not a convenience.