VLASH: Real-Time VLAs via Future-State-Aware Asynchronous inference - Jiaming Tang 03/09/2026

In this talk

1. 1:30

   Why synchronous inference slows robots

   Tang says many VLA systems make the robot wait for each action chunk to finish before requesting the next one. That leaves the GPU unused during movement and can delay responses to changes, such as a dropped object. He cites execution periods around 500 milliseconds as a barrier to fast reactions.

2. 3:45

   Asynchronous inference needs state alignment

   VLASH runs the next inference while the robot carries out its current actions, then switches to the new chunk when it is ready. Tang says this can cut reaction latency by 10 to 20 times, but a direct switch can make movement jump because the prediction was based on an earlier robot state. He describes action blending in Real-Time Chunking as a useful approach that adds about 40% runtime overhead.

3. 10:04

   Train the model to use future robot states

   VLASH estimates the robot state at the time a delayed prediction will be executed, using its current state and the actions already underway. Tang says existing models may underuse robot-state inputs, so fine-tuning shifts robot states and matching target actions by a randomly sampled delay while holding the image input fixed. A shared-observation attention method trains multiple delay offsets together, allowing about five times more trajectories with roughly 20% more computation.

4. 15:54

   Trade action detail for speed and evaluate results

   Tang describes grouping fine-grained actions into larger steps and resampling the trajectory, which can speed physical execution while retaining its path. In tests across simulation and real-world manipulation tasks, he reports that factors of 1.5 to 2 can speed execution without reducing success rates versus synchronous inference in the tested settings. He also reports reaction latency falling from 523 milliseconds to 23 milliseconds on an H100, with an inference rate near 30 Hz on an RTX 5090.

5. 25:06

   Deployment needs data and accurate state estimates

   In Q&A, Tang says deployments typically use 50 to 100 demonstration episodes, collected at normal speed, and that action quantization can make execution faster than the demonstrations. He explains that future robot states can be estimated from current states and planned actions, but future observations are unavailable. If actual states diverge too far from estimates, performance can suffer; Tang says the model tolerates some error.