In the race to build general-purpose humanoid robots, researchers have hit a frustrating wall: the "intelligence vs. speed" trade-off. As we scale up Vision-Language-Action (VLA) models to make robots smarter and better at following instructions, the models get slower. For a robot in the real world, "thinking" for an extra 100ms means acting on a world that has already moved on.
A new paper from researchers at XPENG Robotics and the University of Washington introduces DASH (Decoupled Asynchronous System for Humanoids), a system that breaks this coupling, allowing for "latency-agnostic scaling."
TL;DR
DASH decouples slow, high-level reasoning (System 2) from fast, reactive control (System 1) in VLA models. This allows humanoid robots to use massive 8B+ parameter backbones for superior generalization while maintaining a constant, high-frequency 20Hz control rate, significantly improving success in complex dexterous tasks.
What Problem Does This Paper Address?
Standard VLA models are typically synchronous: the robot captures an image, the entire model (backbone + action head) processes it, and then the robot moves.
As backbones grow from 0.8B to 8B parameters to improve "common sense," their inference latency skyrockets. In a closed-loop system, this creates staleness. By the time a large model finishes "thinking" about an image taken at time $t$, the robot is already at time $t+L$. Executing that action leads to jerky, hesitant, and spatially misaligned movements. This is particularly devastating for dexterous humanoid hands that require high-frequency, contact-rich reactivity.
Methods
The core innovation of DASH is the decoupling of the policy into two concurrent processes:
- System 2 (The Slow Brain): A large, frozen Vision-Language Model (VLM) backbone (like Qwen3-VL) that processes high-level semantics and instructions. It runs at its own natural, slower pace.
- System 1 (The Fast Reflexes): A lightweight flow-matching action head that runs at a fixed, high frequency (20Hz).
Key Technical Components:
- Cross-Timescale Fusion: The action head doesn't just wait for the VLM. It has its own fast ResNet-152 encoder to grab "fresh" visual tokens. It then fuses these with "stale" semantic features cached from the VLM.
- Desynchronized Training: To make the action head robust to this staleness, the authors train it by randomly offsetting the VLM's input time relative to the action head's input. This teaches the model to reconcile old "intent" with new "perception."
- Adaptive Temporal Aggregator: Since inference times can vary slightly, DASH uses a novel fractional interpolation method to smoothly blend overlapping action chunks, preventing the "jitter" common in asynchronous systems.
- LCM Protocol: The system uses the Lightweight Communications and Marshalling (LCM) protocol for non-blocking communication between the "brain" and the "reflexes."
Datasets
- RoboTwin: A 50-task simulation suite for humanoid manipulation. The authors modified it to be asynchronous, charging the model for its inference time in simulated seconds—a much more realistic test than standard "paused-clock" simulations.
- Real-World Humanoid: Evaluation on a physical humanoid with two dexterous hands performing tabletop pick-and-place tasks with various household objects (cola cans, milk cartons, toy cars).
Key Insights and Ideas
- The "Uncertain Tail" of Action Chunks: VLA models predict a "chunk" of future actions. The paper observes that the beginning of a chunk is highly reliable, but the "tail" becomes increasingly uncertain and divergent.
- Frequent Refreshes are Better than Long Thinking: By running the action head at 20Hz, DASH constantly "refreshes" the action plan, staying on the reliable early portion of the predictions and avoiding the oscillatory corrections caused by executing the uncertain tail.
- Fresh Vision is Non-Negotiable: The authors found that simply unfreezing the VLM isn't enough. A dedicated, high-speed vision encoder in the action head is essential to provide the "fresh" evidence needed for reactive control.
Key Results
- Latency-Agnostic Scaling: While a synchronous pipeline's control rate dropped from 9.7 Hz (0.8B model) to 4.1 Hz (8B model), DASH maintained a rock-solid 20 Hz across all model sizes.
- Superior Success Rates: In simulation, DASH outperformed the synchronous baseline at every scale, with the largest gain (+5.9 percentage points) seen at the 4B parameter size.
- Real-World Performance: On a physical humanoid, DASH achieved a 95.8% success rate (23/24 trials) compared to only 66.7% (16/24 trials) for the synchronous system. The DASH-controlled robot was visibly smoother and more reactive to perturbations.
Limitations and Future Work
The authors acknowledge that the tasks evaluated were largely quasi-static (tabletop manipulation). While DASH excels here, the real test will be highly dynamic tasks—like catching a flying object or reacting to a moving person—where the 20Hz refresh rate and fresh vision encoder will be even more critical.
Future work will likely focus on extending this asynchronous framework to long-horizon tasks and even larger "System 2" models that might take seconds to process, further testing the limits of how much "staleness" a reactive system can handle.
Summary updated