Robots run by vision-language-action models lose time at four stages, and a new roundup of 25 techniques for faster VLA robots groups the fixes by where each one saves it. If an object moves mid-reach, the robot’s next action is computed from a camera frame that is already stale.
Some shrink the computation in the network, some produce actions with fewer passes, some let the arm keep going during inference, and some skip network trips. Each paper reports its own test setup, so the figures don’t line up.
Cutting work inside the model
These help when each model call is the bottleneck. APS-VLA trains a selector on demonstrations to keep only useful visual tokens. On two dual-arm platforms, over five real tasks, average inference time dropped from 137.25 to 93.25 ms, with 61% success versus 60% unpruned. ActPrune picks image patches using the prior action chunk and made OpenVLA-OFT 1.61 times faster on LIBERO while keeping 25% of tokens. On π0.5 the gain was just 1.03 times, as the action head dominates there.
EfficientVLA drops repetitive layers and redundant tokens, reaching 1.93 times faster inference on an NVIDIA A40 for a 0.6-point success loss. VLA-Cache needs no retraining: it recycles stored attention data for image regions that barely changed, giving up to 1.7 times faster CUDA inference. QuantVLA shrank two π0.5 modules from 4.27 GB to 1.28 GB, a memory result whose speed benefit depends on hardware. SmolVLA takes the small-model route, with about 450M parameters and 64 visual tokens per image.
Keeping the arm moving
In asynchronous setups the model plans ahead during motion. Real-Time Chunking begins the following prediction early and locks in actions already committed, so the robot progressed faster despite added delay. Training-Time RTC builds that into training; with π0.6 on a remote H100, mean request latency fell from 135 to 108 ms.
SR-VLA predicts upcoming sensor input and starts inference early, trimming π0.5’s wait per query in simulated LIBERO from 128 ms to 48 ms. DASH separates a fast action head from the slow backbone and held 20 Hz for backbones of 0.8B to 8B parameters in simulation, where a synchronous design sank from 9.7 to 4.1 Hz. Path-preserving Time Scaling adjusts arm speed along the path and improved completed-task throughput 28% for π0.5 and 38% for SmolVLA. RAVEL, VLASH and A2C2 target the same lag. For VLASH the authors calculated a worst-case reaction delay of 46.4 ms versus 546.4 ms synchronously, and A2C2 lifted success on a delayed LIBERO Spatial test from 64.4% to 84.2%.
Generating actions in fewer steps
Zora drafts actions by retrieving what happened next in comparable earlier runs, with no second neural model. Software alone gave up to 3.4 times speedup on NVIDIA Thor; the 7.6 times number assumes a proposed hardware accelerator tested in simulation. VSpark’s speculative decoding sped four token-by-token VLAs up by 1.28 to 2.10 times on average.
OpenVLA-OFT outputs a whole block of continuous actions at once, multiplying action throughput by 26 and lowering latency from 239.6 to 72.9 ms in its one-image test. SnapFlow distills a multi-step flow policy into one step, scoring 98.75% on LIBERO against its teacher’s 97.75%. FlashVLA streams overlapping action chunks and sustained 30 Hz or better on one GPU on a real robot.
One caution before touching an action head: the unmodified π0.5 with a single denoising step already ran 3.4 times faster, though success slipped to 96.75%. That baseline is worth trying first.
Reducing round trips to the server
A fast remote GPU can still keep a robot idle through transfer and queueing. VLA-ULAP has a small onboard model bridge the gaps, cutting remote calls by 70 to 72% per successful episode in a real SO-101 task. VLA Handover sent 49 to 77% fewer requests to the server across three simulated LIBERO suites, at a cost of 2 to 4 points of success. LIBERO-Long was harsher: 63.6% against 93.2% for always using the server. Robion focuses on shared servers and met 98% of deadlines for as many as 64 simulated robots on four GPUs.
How to choose a method
First measure where the delay sits: reading the scene, producing a chunk, idling between chunks, or communicating with another machine. A quicker model call, a quicker reaction and a shorter task are separate outcomes, and a method may help one without touching the rest. Check which clock each paper timed before comparing results.
Source: Turing Post