DASH keeps humanoid control fast as models grow
Decoupled Asynchronous System for Humanoids (DASH): Latency-Agnostic Scaling of Vision-Language-Action Models for Humanoid Manipulation · arXiv:2611.11854
In plain words
A robot can keep moving while a large vision-and-language model works out what to do next. By the time the answer arrives, the scene may have changed.
DASH is a robot control system that lets a fast action process keep working while the slower model updates its guidance, and it gives the fast process fresh camera images.
In a simulated test, DASH kept the robot’s control rate steady as the models grew larger. Its reported task success was also higher than synchronous control at each tested size, though the study’s results cover specific simulated tasks and tabletop trials.
The research
Humanoid manipulation needs both instruction understanding and timely corrections: waiting for a large model can leave a robot acting on an outdated scene.
DASH’s results make it plausible to scale scene understanding without sacrificing action updates in the tested tasks, while showing that training for delayed context matters too.
DASH separates slow scene understanding from fast robot actions, then trains the action process to pair fresh images with older model context.
In a 50-task RoboTwin simulation, researchers charged model delay as simulated time while robots continued their previous actions. Across four model sizes, DASH held 20 control updates per second; synchronous control fell to 4.1 at 8B. Reported task success was higher for DASH at every size, peaking at 69.5% at 4B.
The simulation covered 50 RoboTwin tasks, and the real-robot comparison involved 24 tabletop trials per setup. That comparison also changed the execution pipeline and chunk aggregator; the study describes its tasks as quasi-static and leaves dynamic and long-horizon tasks for future work.
What the evidence shows
- Separate understanding from action
DASH, the Decoupled Asynchronous System for Humanoids, is a robot manipulation system that separates a slower vision-and-language model from a faster action process. The question is whether a robot can retain the larger model’s scene understanding without waiting for it before each action.
DASH lets the action process continue using fresh images and the latest available model context, even when that context is delayed. F1, F2.
- Keep control moving as models grow
The comparison shows that splitting the fast action path from the model kept control rate independent of backbone size in this simulation.
Reported task success also favored DASH at each tested size, but the comparison changes runtime, training alignment, and aggregation, so it does not identify which change explains that difference.
- Test the role of training
The success comparison needs a closer look because the main systems differ in more than their runtime: DASH also uses training that exposes the action process to delayed model context. Within the asynchronous runtime, aligned training produced 62.3% reported success at 4B, compared with 69.5% after desynchronized training.
That contrast points to the training recipe as part of the result, though it applies within the split runtime and does not isolate the cause of the full DASH-versus-synchronous difference.
- Check the visual input recipe
This comparison tests a combined recipe, not the encoder by itself: freezing the backbone and adding fresh ResNet vision achieved the highest reported success among the three conditions.
Because the recipe comparison also changes whether the backbone is trained further, it supports that combination in the 4B asynchronous simulation, not a single ingredient.
- Carry the test onto a robot
DASH succeeded in 23 of 24 language-guided tabletop trials, compared with 16 of 24 for synchronous inference with fixed-offset chunk aggregation.
This real-humanoid comparison extends the evidence beyond simulation, but it changes both execution pipeline and aggregation, and the study describes its evaluated tasks as quasi-static.
Deep
- Original paper authors
- Bingzhang Wang, Nathan Zhao, Jialeng Ni, Jiajin Zheng, Yin Zhang
- Submitted to arXiv
- License
- arXiv non-exclusive