A robot-control system called DualWAM separated broad visual planning from fast local correction, cutting the action path to 63 milliseconds while retaining a slower world model in the background. Across 20 unseen manipulation tasks on Franka and Galbot robots, it improved average success by 4.5 percentage points over the strongest evaluated baseline.
World action models generate robot actions while predicting how the scene will change. Future-frame prediction can provide useful physical context, but it is expensive. Existing systems often amortize that cost by issuing long chunks of actions, which makes them slower to react when contact, object motion or the robot's actual path differs from the plan.
The model begins with noise representing possible future views and actions, then refines it into a predicted scene and movement sequence. DualWAM splits this denoising process between two systems. System 2 periodically handles the high-noise portion over a 32-frame horizon, using global and wrist cameras to establish a world-action plan. System 1 takes an aligned eight-frame slice of that intermediate state and finishes the low-noise portion using fresh wrist-camera observations. The global planner runs asynchronously, so several local corrections can reuse its latest valid state instead of waiting for a new plan.
Real-robot trials separate success from latency
The authors evaluated ten tasks on a Franka FR3 and ten on a Galbot G1. Each policy received ten trials per task, giving 100 rollouts for each robot. A task counted as zero-shot only when its target objects and scene were absent from robot training tasks and the complete skill-object-environment combination was unseen. There was no task-specific demonstration or fine-tuning.
DualWAM completed 54% of Franka trials and 45% of Galbot trials. The strongest comparison, Cosmos3-Nano-Policy, reached 49% and 41%. Successful DualWAM episodes also finished faster on average: 21.76 seconds versus 24.88 seconds on Franka, and 48.67 seconds versus 53.11 seconds on Galbot. These completion times exclude policy inference and measure executed control steps divided by control frequency.
The critical-path comparison is sharper. DualWAM required 63 milliseconds, while the optimized Cosmos3-Nano-Policy path required 1,043 milliseconds, a 16.6-fold difference. The paper says both received the same system and implementation optimizations for this latency test. The deployed DualWAM used three RTX 4090 GPUs, so the result demonstrates response time on a multi-GPU research setup rather than an embedded robot computer.
Separating the effects of prediction and scheduling
An architecture ablation on Franka helps separate the causes. System 2 alone achieved 49% success at 1,043 milliseconds. Combining the global planner with DualWAM's local world-action refiner raised success to 54% and reduced the path to 63 milliseconds. Replacing that local refiner with an action-only diffusion transformer kept latency near 59 milliseconds but reduced success to 7%. A Fast-WAM local model that used video-action co-training without predicting future video at inference reached 14% at 62 milliseconds.
A second experiment kept the two models and denoising schedule fixed while changing their coordination. On Galbot, synchronous execution reached 39% success and 850 milliseconds. Reusing the latest global state asynchronously raised success to 45% and reduced the action path to 63 milliseconds. This supports the scheduling claim within the authors' setup, although it does not establish how far a global plan can become stale before local corrections fail.
The split also reduces what a cloud planner must send to an edge controller. At the same 0.469 Hz synchronization rate, DualWAM sent 0.423 MB downstream per update, compared with 44.2 MB for AHA-WAM's layer-wise video context. Average downlink bandwidth fell from 20.7 to 0.198 MB/s. That is a payload comparison on the study's representation and network assumptions, not a measurement of public-cloud service latency.
Training remains expensive. Both systems were trained for 10,000 iterations with a global batch of 1,024 across 64 NVIDIA H200 GPUs. The authors also report that adding 1,800 hours of egocentric video to the global system and 1,200 hours of wrist-centric demonstrations to the local system raised Galbot success from 45% to 59%. Because those data additions are bundled, the 14-point change should not be attributed to the architecture alone.
Sources
- DualWAM paper record and abstract, submitted September 21, 2026
- Full paper with architecture, real-robot trials and ablations
- Representative robot-arm image by Zoe-Rose Herbert for the UK Department for Science, Innovation and Technology, CC BY 2.0
- Lead image: a representative laboratory robot arm, not either robot used in the study.