Humanoids stop waiting for the big model: DASH holds 20 Hz from 0.8B to 8B backbones
Synopsis
In humanoid manipulation, a large vision-language-action backbone is slow, so synchronous execution acts on stale observations and control frequency falls as the backbone grows. DASH splits the policy into a slow vision-language model and a fast flow-matching action head running as three concurrent processes over LCM, so the control rate is set by the action head alone. On a RoboTwin benchmark that charges inference latency in simulated time, DASH holds 20 Hz from 0.8B to 8B backbones while the synchronous rate falls from 9.7 Hz to 4.1 Hz, with higher success at every size; on a real dexterous humanoid, pick-and-place succeeds in 23 of 24 trials versus 16 of 24 for the synchronous pipeline.
Interpretation
DASH splits the vision-language-action policy on a humanoid into two modules on independent clocks: a slow vision-language model for semantics and a fast flow-matching action head that emits actions, so control frequency no longer drops as the backbone grows. Such models previously ran synchronously end to end, and an action became usable only after the backbone finished, so a larger backbone meant longer latency and a lower control rate. On a RoboTwin benchmark that charges inference latency in simulated time, DASH holds 20 Hz at 0.8B, 2B, 4B and 8B backbones, while the synchronous control rate falls from 9.7 Hz to 7.6, 5.5 and 4.1 Hz.
The action head combines stale semantics from the frozen backbone with fresh vision from its own ResNet encoder, and training randomly offsets the two inputs in time so the staleness met at deployment has already been seen. Synchronous training only sees inputs aligned at a single timestamp, whereas at deployment the action head faces semantics one backbone cycle old. On the 50-task RoboTwin suite with 250 episodes per configuration, DASH has higher success than synchronous inference at all four backbone sizes, peaking at 69.5% with 4B, 5.9 percentage points above synchronous; the same asynchronous runtime trained on aligned data sits slightly below synchronous inference.
Overlapping action chunks are aligned by a fractional offset measured from actual arrival intervals and resampled by linear interpolation, so the higher chunk density of asynchronous inference becomes finer motion rather than jitter. Synchronous inference has a fixed period, so the chunk offset is a known integer; under asynchrony the inference duration varies per chunk and the offset becomes a time-varying real number, and rounding it over-advances execution. On a real humanoid with two dexterous hands doing language-specified tabletop pick-and-place, DASH with the adaptive aggregator succeeds in 23 of 24 trials, versus 16 of 24 for synchronous inference with a fixed-offset aggregator.
Perspective
Evaluation is limited to quasi-static tabletop pick-and-place: simulation uses the 50-task RoboTwin suite, and the real platform is a humanoid with two dexterous hands executing language-specified pick-and-place; the authors explicitly name dynamic and long-horizon tasks as the next evaluation step.
The authors state that the evaluated tasks are quasi-static and expect the constant control rate and fresh perception to matter most on dynamic and long-horizon tasks, which have not been evaluated; part of the simulated suite is out of distribution relative to the training demonstrations, which keeps absolute success rates moderate.