Decoupled Asynchronous System for Humanoids (DASH): Latency-Agnostic Scaling of Vision-Language-Action Models for Humanoid Manipulation
Decoupled Asynchronous System for Humanoids (DASH): Latency-Agnostic Scaling of Vision-Language-Action Models for Humanoid Manipulation
This paper introduces DASH, a Decoupled Asynchronous System for Humanoids designed to scale vision-language-action (VLA) models without compromising control latency. By splitting the policy into a slow vision-language backbone and a fast action head running concurrently via LCM, DASH maintains a stable 20 Hz control rate regardless of backbone size (tested from 0.8B to 8B parameters), whereas synchronous pipelines degrade significantly as model size increases. The system fuses frozen semantic information with fresh visual inputs and adapts to variable inference latency through desynchronized training and temporal aggregation. Evaluations on a simulated RoboTwin benchmark and a real humanoid robot with dexterous hands demonstrate that DASH sustains high control frequencies while achieving higher pick-and-place success rates compared to synchronous baselines.
The authors propose DASH, a system that decouples vision-language processing from control loops to maintain high-frequency actuation in humanoid robots despite large model backbones.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
control rate
20
View original evidence
frozen backbone with fresh vision from its own ResNet encoder, desynchronized training exposes it to the staleness it meets at deployment, and an adaptive temporal aggregator aligns overlapping action chunks under variable inference latency. On an asynchronous RoboTwin benchmark that charges inference latency in simulated time, DASH holds 20 Hz control rate from 0.8B to 8B backbones while the synchronous rate falls f
Open source S4
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
frozen backbone with fresh vision from its own ResNet encoder, desynchronized training exposes it to the staleness it meets at deployment, and an adaptive temporal aggregator aligns overlapping action chunks under variable inference latency. On an asynchronous RoboTwin benchmark that charges inference latency in simulated time, DASH holds 20 Hz control rate from 0.8B to 8B backbones while the synchronous rate falls f
Open source S4
rom 9.7 to 4.1 Hz, with higher success at every size. On a real humanoid with dexterous hands, DASH achieves clearly higher pick-and-place success than the synchronous pipeline.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org