CVPR 2026 VIRAL

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

Venue: CVPR 2026 Category: Humanoid Loco-Manipulation / Sim-to-Real Trend tag: Trend 2 (humanoid sub-cluster) Affiliations: NVIDIA + CMU + UC Berkeley + CUHK

Approach diagram

flowchart LR
  SIM["simulator with privileged state"] --> TEACHER["RL teacher<br/>privileged-state policy"]
  TEACHER --> ROLL["expert rollouts"]
  ROLL --> STUDENT["student policy<br/>vision-only"]
  TILED["tiled-rendering vision pipeline"] --> STUDENT
  STUDENT --> DEPLOY["zero-shot Unitree G1<br/>54 cycles"]

Loading

Problem

Humanoid loco-manipulation in the real world is bottlenecked by the sim-to-real gap on visual observations. RL policies trained on rendered images often fail to transfer; policies trained on raw images do not learn fast enough.

Method

Standard teacher-student with two scaling refinements:

• Privileged-state RL teacher — trained in simulation with full state access, using a delta action space and reference state initialization (RSI) to learn long-horizon loco-manipulation; produces strong reference rollouts.
• Tiled-rendering student — RGB-based student policy distilled from the teacher via large-scale tiled-rendering simulation, trained with a mixture of online DAgger and behavior cloning, which lets training run with high throughput and high diversity.
• Sim-to-real transfer combines large-scale visual domain randomization (lighting, materials, camera parameters, image quality, sensor delays) with real-to-sim alignment of the dexterous hands and cameras.
• The student is the only model deployed; the teacher exists only to provide rollouts.

A central empirical finding is that compute scale is decisive: scaling simulation to tens of GPUs (up to 64) makes teacher and student training reliable, while low-compute regimes often fail.

Results

54 zero-shot cycles on Unitree G1 humanoid — substantial real-world deployment evidence on a commercially available humanoid platform.

Significance

VIRAL is the infrastructure half of CVPR 2026's humanoid cluster. The companion paper Open-Sim-to-Real (same lab affiliations) is the methods half, attacking articulated-object loco-manipulation. Together they push humanoid sim-to-real to a level where commercially available platforms (Unitree G1) can be trained at scale.

Links

• arXiv: 2511.15200
• Project: viral-humanoid.github.io

Related pages

• Open-Sim-to-Real (companion paper)
• WholeBodyVLA · Visual Imitation → Humanoid
• CVPR 2026 survey

← Back to CVPR-2026