Decoupled Asynchronous System for Humanoids (DASH): Latency-Agnostic Scaling of Vision-Language-Action Models for Humanoid Manipulation

Bingzhang Wang, Nathan Zhao, Jialeng Ni, Jiajin Zheng, Yin Zhang, Kehua Chen, Jian Yao

November 2026

TL;DR DASH runs a VLA's vision-language backbone and its flow-matching action head as separate asynchronous processes, so the control rate no longer depends on backbone size. It holds 20 Hz from 0.8B to 8B backbones (synchronous inference drops from 9.7 to 4.1 Hz) and succeeds in 23/24 real humanoid pick-and-place trials against 16/24.

What they do

In a standard VLA the vision-language backbone sits inside the control loop: while it computes, the robot keeps moving, so a bigger backbone means staler actions. DASH splits the policy into a slow VLM (System 2) and a fast flow-matching action head (System 1) that run on independent clocks, and trains the head to cope with the stale VLM context this creates. It is evaluated in a RoboTwin simulator modified to charge inference latency in simulated time, and on a real humanoid with two dexterous hands.

How it works

Three processes over LCM. Environment, VLM server and action server publish and subscribe without blocking; each reads the latest message. The control rate equals the action-head rate.
Cross-timescale fusion. The VLM (Qwen3-VL) is frozen. A 12-block DiT head cross-attends to cached hidden states from four VLM depths plus fresh tokens from its own ResNet-152 image encoder.
Desynchronized training. The VLM input is drawn from a frame a random number of ticks earlier than the action-head input, up to the worst delay measured at deployment. No explicit staleness input is added.
Adaptive temporal aggregation. Overlapping action chunks are aligned by the measured arrival interval, with the fractional offset handled by linear interpolation and carried over between chunks.

Results

Latency on an RTX 4090: the VLM takes 53 to 193 ms from 0.8B to 8B, so synchronous control falls from 9.7 to 4.1 Hz. The DASH action path stays at 50 ms (20 Hz), 2.1x to 4.9x faster.
RoboTwin, 50 tasks, 250 episodes per configuration: DASH beats synchronous inference at every size by 1.0 to 5.9 points, peaking at 69.5% with the 4B backbone. The asynchronous runtime with aligned training sits slightly below synchronous; desynchronized training adds 3.2 to 7.2 points.
Ablation at 4B: frozen backbone without a head encoder 55.0%, full fine-tune without an encoder 62.6%, frozen plus ResNet-152 69.5%.
Real humanoid pick and place: 23/24 for DASH with the adaptive aggregator against 16/24 for synchronous inference with a fixed-offset aggregator.

Limitations

The evaluated tasks are quasi-static; the authors leave dynamic and long-horizon tasks for future work. The real-robot comparison is 24 pick-and-place trials per configuration, and it changes the runtime and the aggregator together. Part of the simulation suite is unseen in training, which keeps absolute success moderate.

Figures

From Figure 3: DASH architecture. The policy is split into a slow VLM (System 2) and a fast action head (System 1), d... — From Figure 3: DASH architecture. The policy is split into a slow VLM (System 2) and a fast action head (System 1), deployed as three concurrent processes (environment, VLM server, and action server) coupled by a decentralized, non-blocking latest-value bus (LCM). The VLM runs off the control loop, so the control rate is set by the action head alone.

From Figure 4: Desynchronized training data. The two visual channels are sampled at different times so the training i... — From Figure 4: Desynchronized training data. The two visual channels are sampled at different times so the training inputs reproduce the staleness met at deployment, while the supervision target remains anchored at the action head’s timestamp.

From Figure 7: Matched simulation episodes, synchronous inference (top) vs. DASH (bottom). Frames are sampled at equa... — From Figure 7: Matched simulation episodes, synchronous inference (top) vs. DASH (bottom). Frames are sampled at equal intervals of simulated time up to the completion time of the successful run.

From Figure 8: DASH on the real humanoid: three of the language-conditioned pick-and-place instructions, six frames e... — From Figure 8: DASH on the real humanoid: three of the language-conditioned pick-and-place instructions, six frames each at equal intervals of wall-clock time.

Abstract

Vision-language-action (VLA) models pair a vision-language backbone with an action head, but in closed-loop control the backbone sits on the control-critical path: while it computes, the robot keeps moving, so each executed action is conditioned on an increasingly stale observation. A larger backbone therefore directly blocks reactivity, especially in humanoid dexterous manipulation. We present DASH, a Decoupled Asynchronous System for Humanoids that removes this coupling. DASH splits the policy into a slow vision-language model (System 2) and a fast flow-matching action head (System 1) that run as concurrent processes over the LCM communication protocol, so the control rate is set by the action head alone, independent of backbone size. The action head fuses stale semantics from a 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 from 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.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong