DeltaWAM: Delta World Action Models for Bimanual Manipulation

arXiv · Han Yan, Zishang Xiang, Haokai Jiang, Zeyu Zhang, Qilin Wang, Weiyu Guo, Yandong Guo, Boxin Shi, Hao Tang · 2026-09-27

This item is no longer on the current shelf. Its page remains available, and editorial corrections may be made. See what is featured now.

DeltaWAM slashes video-expert cost and boosts bimanual manipulation success by predicting visual deltas instead of full frames.

What it does

DeltaWAM extends World-Action Models (WAMs) by jointly predicting visual deltas and robot actions, using three architectural variants that share representations differently. It introduces Streaming Delta Memory (SDM), a cache that stores an anchor context and updates it with compact observed deltas, eliminating the need to run a heavy video expert on every frame. On the RoboTwin benchmark, DeltaWAM with SDM raises average success from 81.3 % to 85.4 % in a clean setting and from 75.8 % to 83.9 % under visual randomization. Training FLOPs drop by 17.78-23.77 %, while SDM cuts one-step inference latency by 36.57 % and FLOPs by 31.55 %. Real-world tests confirm the highest overall success rate and normalized progress among the compared policies.

Why it matters

Bimanual manipulation demands fast, reliable action generation; processing full video frames at each step is a bottleneck. By focusing on visual deltas, DeltaWAM reduces computational load without sacrificing accuracy, enabling tighter control loops and more scalable deployment on real robots.

How it applies

Robotics teams can replace dense video-expert pipelines with DeltaWAM-SDM to achieve lower latency and power consumption, crucial for edge devices. The three architecture options let engineers trade off representation sharing versus compute based on hardware constraints.

The short version

What if your robot could skip watching the whole movie and just read the subtitles?

My take

I’d start swapping any full-frame prediction module with a delta-based counterpart, especially when the scene changes slowly relative to the robot’s action frequency. The SDM cache is a pragmatic way to keep context without re-encoding the entire visual stream, and the three architecture flavors let you match the model to your robot’s compute envelope.

How it connects

• Delta-focused perception mirrors the broader trend of sparse, event-based sensing in robotics.
• Efficient world-action models are a stepping stone toward truly real-time, multimodal agents that can reason and act on the fly.

Bottom line

Adopt visual-delta prediction and streaming memory to make bimanual agents faster, cheaper, and more reliable.

Takeaways

• Predicting visual deltas instead of full frames cuts training FLOPs by up to 23.77 %.
• Streaming Delta Memory reduces inference latency by 36.57 % and FLOPs by 31.55 %.
• Success rates improve to 85.4 % (clean) and 83.9 % (visual randomization) on RoboTwin.
• Three architectural variants provide flexibility for different hardware budgets.
• Real-world experiments confirm DeltaWAM’s superiority over prior WAMs.

Topics: robotics, bimanual manipulation, world-action models, efficiency, streaming memory