What Is VLA-Cache?

VLA-Cache is introduced in VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (arXiv 2502.02175) by Siyu Xu, Yunke Wang, Chenghao Xia, Dihao Zhu, Tao Huang and Chang Xu, first submitted on 4 February 2025 and accepted to NeurIPS 2025. The abstract starts from a cost problem: VLA models reason well over vision and language, but their computational cost is a challenge for real-time robotic control, where rapid decision-making is essential. VLA-Cache attacks that cost at inference time, without any retraining.

VLA-Cache is third-party research. Firsthand is not part of it or affiliated with its authors, and makes no claim that any model was trained on Firsthand data. What the site states is custom egocentric collection of real human demonstrations, fast and on demand, to a written spec: 3D hand pose at 21 joints per hand, calibrated stereo depth and 6-DoF IMUs at 200 Hz, delivered in RLDS, LeRobot format or WebDataset with a buyer-owned commercial license. The only language data the site states is one task instruction per episode. Whether a VLA-Cache-style speed-up changes what data a program needs is Not stated in the sources read.

Per the abstract, VLA models enable direct action generation from visual perception and language instructions in an end-to-end manner, but their substantial computational cost poses a challenge for real-time robotic control. The GitHub README puts the cause in the data: the high frame rate and spatial complexity of robotic control make VLA inference computationally expensive.

The authors exploit the temporal continuity of robotic manipulation. VLA-Cache identifies visual tokens that changed minimally between adjacent frames and reuses their cached key-value representations, so the model avoids redundant computation on parts of the scene that did not move.

To maintain action precision, it selectively re-computes task-relevant tokens that are environmentally sensitive, which keeps critical visual information accurate. A third part, a layer adaptive token reusing strategy, adjusts the reuse ratio dynamically based on attention concentration across decoder layers, prioritising critical tokens for recomputation.

No. The abstract calls it a training-free inference acceleration method, and the README describes it as a training-free, plug-and-play solution for accelerating vision-language-action models. It changes how an existing model runs, not how it is trained, so it needs no new demonstration data of its own.

The abstract reports experiments on two simulation platforms, LIBERO and SIMPLER, and a real-world robotic system. It states up to a 1.7x speedup in CUDA latency and a 15% increase in control frequency, with negligible loss on task success rate.

The project page adds the setups: LIBERO comparisons use OpenVLA as the baseline model, SIMPLER comparisons use CogACT, and the real-world tests use a Kinova Jaco2 arm that collected demonstrations to fine-tune OpenVLA across four tasks. These are the authors’ results on their own setups. Firsthand does not reproduce them or quote per-suite scores.

The official repository, siyuhsu/vla-cache, is released under the Apache License 2.0. Its README lists code for OpenVLA and for OpenVLA-OFT, with evaluation scripts that run the LIBERO suites with the cache switched on or off so the two can be compared. The README says it builds on OpenVLA, OpenVLA-OFT and Hugging Face Transformers.

The site’s TinyVLA entry covers compact VLA models that start from a fast multimodal backbone and need no robot-data pretraining stage, and its BitVLA entry covers a model whose weights are ternary values. Both change the model itself.

VLA-Cache changes neither the parameter count nor the numeric precision. It leaves the model as it is and removes repeated work between frames while it runs. All three are third-party efficiency ideas, and Firsthand ranks none of them.

This entry draws on the arXiv abstract page, the official GitHub repository and the official project page. None of them describes human egocentric video as part of VLA-Cache’s evaluation, and none states how stereo depth, hand pose or IMU data would interact with token reuse, so those are Not stated.

When a program needs real human egocentric demonstrations of tasks, environments or objects that public data does not cover, specified in writing. Firsthand does not build or train models, and the language it delivers is one task instruction per episode.

Confirm the spec on a Spec Pilot, a small fast first batch. Where coverage is deep, the median from signed spec to first validated batch is 48 hours; where it is thin, three to four weeks. Billing is per validated hour.

What is VLA-Cache?

A published training-free method (arXiv 2502.02175) that speeds up vision-language-action models by caching and reusing the key-value representations of visual tokens that barely change between frames.

What does adaptive token caching mean in VLA-Cache?

Static visual tokens are reused across frames, task-relevant tokens that are sensitive to the environment are recomputed, and a layer adaptive strategy sets the reuse ratio from attention concentration across decoder layers.

Does VLA-Cache require retraining the model?

No. The authors describe it as training-free and plug-and-play, and its official code runs on OpenVLA and OpenVLA-OFT.

What efficiency did the authors report?

The abstract states up to a 1.7x speedup in CUDA latency and a 15% increase in control frequency, with negligible loss on task success rate. These are the authors’ results on their own setups.

Does Firsthand data train VLA-Cache?

No claim is made. VLA-Cache is training-free, Firsthand is not affiliated with its authors, and the site does not state that any model was trained on Firsthand data.

Can Firsthand collect egocentric data for a VLA program?

The site states custom egocentric collection to a written spec, fast and on demand, with 21-joint 3D hand pose, calibrated stereo depth, 200 Hz IMUs and one task instruction per episode, confirmed on a Spec Pilot.