Robot lag turns out to be a scheduling problem: VLASH plans the next move mid-motion

A team from MIT, Nvidia, Caltech, UC Berkeley, UC San Diego and Tsinghua has published a robot control scheme called VLASH that computes the next batch of actions while the current batch is still executing, and reports cutting maximum reaction latency by as much as 11.8 times without additional computing power ^[7][6][1]. For anyone whose vision-language-action pilot moves in a reach-pause-adjust-pause rhythm, that locates the defect in the inference loop rather than in the GPU budget ^[2].

The conventional arrangement is serial. The robot finishes one set of instructions, then waits while the model computes the next, which produces the stop-and-go cadence ^[2]. According to the researchers, that lag is the main obstacle to using VLA-controlled robots for work that demands continuous, real-time interaction ^[3]. VLA models themselves take camera images, a human instruction and the robot's state, and emit movements ^[10].

VLASH takes the robot's current position plus the movements it has already scheduled, estimates the state the robot will be in when the current chunk ends, and plans the next chunk from that projected state ^[11][4]. Because the estimate comes from motion that is already committed, it needs no separate prediction model and no extra inference pass at runtime ^[13]. It also does not model the surrounding scene: the authors note that predicting the environment with a world model costs more compute, which is the wrong trade when reaction time is the binding constraint, though they say the two approaches could eventually be combined ^[12].

Reported outcomes: some tasks finished 1.5 to 2 times faster while retaining most or all accuracy ^[5]. The gap between the two headline numbers is the interesting part. The worst-case latency improvement is roughly six to eight times larger than the end-to-end task speedup ^[20], which is what you get when you remove the tail of a wait rather than the whole of it.

Testing used two VLA models on two robot platforms with a laptop carrying an Nvidia RTX 5090, covering pick-and-place, stacking and sorting at 20 trials per method, plus fast-reaction challenges including table tennis and Whac-a-Mole; latency was measured separately across various processing units ^[14]. VLASH does not predict where a moving target is going. It re-reads observations 15 to 30 times per second ^[17], which is a fresh look every 33 to 67 milliseconds ^[22].

Provenance matters here. An earlier version of the paper claimed more than a 30x speedup; the authors told Live Science that those runs used longer action sequences and slower hardware, while the newer experiments used shorter, more common sequences and more powerful GPUs, and that "The 11.8x figure may better represent recent practical settings" ^[8][9]. The revised figure is under 40 percent of the earlier one ^[21], a demonstration of how much these multipliers depend on chunk length and silicon. The team also reorganised existing training data and measured one benchmark training step running 3.26 times faster at comparable accuracy, while cautioning that total training does not become 3.26 times cheaper because cost also depends on model, hardware, data and step count ^[15][16].

Speed is not safety. Roshni Lulla of the Institute for Humane Robotics, who was not involved, told Live Science that "Reacting sooner is a real benefit, but the case for improved safety is unclear to me" ^[18].

Watch the IROS presentation and whether the 11.8x holds on shorter action chunks and slower accelerators ^[7][8], and watch which manipulation tasks with continuous motion actually move first, which the authors expect to include manufacturing settings where objects shift and plans change ^[19].