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.
Figures & tables
Fig. 2: (a) Uncertainty grows along the prediction horizon: the early portion of an action chunk is reliable, while its distant tail admits increasingly different future actions. (b) Frequent chunk updates curb divergence: synchronous inference executes the long, uncertain tail of each chunk, so a new prediction can trigger a large correction; DASH refreshes chunks often and stays on the reliable early portion.
Fig. 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 and publishes a context (the hidden states of the routed VLM layers) that the action head reads with latest-value semantics, so the control rate is set by the action head alone.
Fig. 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.
Fig. 5: Adaptive interpolation for temporal aggregation. The fractional offset measured from the true chunk arrival interval is split into a whole-tick shift and a linear interpolation between adjacent entries, with the residual fractions accumulated so the alignment tracks elapsed execution time.
Latency (ms)
Control rate (Hz)
Success rate (%)
Backbone
Sync
DASH
Sync
DASH
Speedup
Sync
Async (w/o desynced training)
DASH
Δ
0.8B
103
50
9.7
20
2.1 ×
64.5
61.8
65.5
+ 1.0
2B
132
50
7.6
20
2.6 ×
66.4
65.5
69.1
+ 2.7
4B
183
50
5.5
20
3.6 ×
63.6
62.3
69.5
+ 5.9
8B
243
50
4.1
20
4.9 ×
62.7
62.7
65.9
+ 3.2
TABLE I: Backbone scaling, synchronous inference vs. DASH (50 tasks, 250 episodes). Async (w/o desynced training): the DASH runtime with aligned training. Δ = DASH − Sync (pp).
Component
Latency (ms)
VLM, Qwen3.5-0.8B
53
VLM, Qwen3-VL-2B
82
VLM, Qwen3-VL-4B
133
VLM, Qwen3-VL-8B
193
Image preprocessing
15
Action head (ResNet-152 + 4-step flow-matching DiT)
Fig. 6: Success rate by backbone size (Table I ): synchronous inference, Async (w/o desynced training), and DASH. DASH is higher than synchronous inference at every size ( +1.0/+2.7/+5.9/+3.2 pp) while its action path stays at 50 ms.
Fig. 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; done marks completion, and still running marks the run that later times out.
Fig. 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 (time stamps in the corners).
Method
Success
SR (%)
Sync, fixed-offset aggregator
16/24
66.7
DASH, adaptive aggregator (ours)
23/24
95.8
TABLE III: Real-robot pick and place ( 24 trials per configuration).