DeltaWAM cuts robot control compute by predicting only frame changes

single source· 1 articles · confidence: high · first seen 2026-09-22 20:00 UTC

What this means for you

If you train video-pretrained robot policies, the transferable idea is predicting frame deltas instead of full frames; the 36.6% saving is one-step inference latency on a simulated benchmark. For everyone else, nothing to act on — code has been posted, but no weights, pricing or API.

DeltaWAM, a robot-control model, removes most of the video-model compute from two-armed manipulation. World-action models borrow a pretrained video generator to predict the next frames and the actions to take; the existing ones re-model every pixel of every frame. DeltaWAM predicts only the deltas between frames and caches an anchor context. Its authors report training compute down 17.8–23.8%, one-step inference latency down 36.6%, and RoboTwin simulator success of 85.4% against 81.3% for Fast-WAM, rising from 75.8% to 83.9% under visual randomisation. It is a preprint, and no evaluation date is given for those numbers.

Key facts

  • ·On the RoboTwin simulator, DeltaWAM with Streaming Delta Memory reports average success of 85.4% versus 81.3% for Fast-WAM in the clean setting, and 83.9% versus 75.8% under visual randomisation. source
  • ·The three proposed architectures reduce training FLOPs by 17.78–23.77%. source
  • ·Streaming Delta Memory reduces one-step inference latency by 36.57% and one-step inference FLOPs by 31.55%. source
  • ·Real-world evaluations are reported to give the highest overall success rate and normalised progress among the policies evaluated, with no baseline names or task details given in the abstract. source
  • ·Code is posted at github.com/AIGeeksGroup/DeltaWAM and a project page at aigeeksgroup.github.io/DeltaWAM. source
  • ·The work is a preprint, arXiv 2609.28811, posted on 22 September 2026. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire