AI

Robust-WAM Keeps Video Pretraining and Adds Semantic Foresight

Robb Harlan 5 min read

World-action models that ride video generation pretraining are strong on dynamics and soft on appearance shifts. Flip the design into a pure semantic latent and you gain lighting robustness — but you lose the web-scale VAE video prior. Robust-WAM (arXiv 2608.05903) is a post-training recipe that keeps the VAE video path and injects semantic foresight into the action stream.

Robust-WAM motivation: VAE-only vs semantic-only vs combined
Motivation: VAE pretraining vs semantic robustness vs Robust-WAM. Source: arXiv:2608.05903 / project page.

The trick

On top of an existing video-generation WAM:

  • Leave the Video DiT denoising VAE latents alone (pretraining stays)
  • Prepend learnable query tokens to the action DiT stream
  • Align each query’s output to the frozen DINOv3 CLS embedding of the matching future ground-truth frame
  • Reuse the positional encoding of the action step that reaches that future frame

Teacher and alignment head are training-only. At inference, only the queries remain.

Robust-WAM architecture with query tokens and DINOv3 alignment
Architecture: video branch retained; action stream gets semantic query alignment. Source: arXiv:2608.05903.

Numbers that matter

From the paper and project page (success rates %):

LIBERO-Plus (train on standard LIBERO, test under camera/lighting/background/layout/noise axes):

Base WAMClean LIBEROLIBERO-PlusWith Robust-WAM (Plus)
FastWAM97.6 → 97.949.758.9 (+9.2)
GE-Act96.5 → 97.378.080.9 (+2.9)

On FastWAM, large Plus gains include sensor noise (+18.9), camera (+12.4), lighting (+11.3), and background (+10.3).

RoboTwin clean→random (LingBot-VA unified WAM): random success 29.8% → 34.4% (+4.6) with clean held at 81.2%.

Real Franka (train under white light; evaluate purple/cyan only): GE-Act OOD average 57.3% → 80.0% with Robust-WAM; InD roughly preserved (80.0% → 82.7% on the reported averages).

Real-robot tasks under standard purple and cyan lighting
Franka tabletop tasks under standard vs purple/cyan illumination. Source: arXiv:2608.05903.

A Human’s Take

Lighting OOD is where a lot of “it worked in the demo” dies. Robust-WAM’s bet is correct for operators: don’t throw away the video prior, post-train the action head to care about scene semantics. I’ll trust this more when third parties re-run the Franka purple-light protocol. Until then, the LIBERO-Plus jumps on FastWAM are the clearest receipt in the paper.

Sources