1. What robot foundation models and VLAs are (the short version)

• A VLA “sees, listens and moves”: give it camera images, an instruction in words (for example “put the red cup on the plate”) and the robot’s state, and it returns several steps of actions. What they share is that they build on a vision-language model (VLM) pre-trained on web images and textSourced[Ref. 2]
• How actions are output is the crux of the design: there are methods that cut each dimension into 256 bins and emit “tokens” (RT-2, OpenVLA), a method that compresses them into fewer tokens (FAST), and methods that generate smooth continuous values (π0, GR00T N1)Sourced[Ref. 2, 5, 8, 7, 11]. For dexterous work at high frequencies such as 50 Hz, this choice matters
• 97% on a benchmark is not 97% on the shop floor: even on the same LIBERO benchmark, the number changes with inputs and fine-tuning set-upSourced[Ref. 6]. There are also reports (preprints) that performance falls sharply when conditions are changed a littleSourced[Ref. 16, 17]. Read a success rate together with the robot, task, number of trials, whether it was fine-tuned and whether it was autonomous

2. The “foundation model” analogy — learning from the experience of many robots

The AI that moves a robot has usually been built afresh for each machine and each job: collect data and train for one task on one robot. The foundation-model idea changes this. First learn general abilities from diverse, large-scale data (pre-training), then adjust with data from your own robot and your own task (fine-tuning).

For example, Open X-Embodiment pooled data on 527 skills (160,266 tasks) collected by 21 institutions on 22 kinds of robot into a standard format, and reports that a model trained on it showed “positive transfer” from the experience of other robots [Ref. 3]Sourced. Octo is an open model trained on 800,000 trajectories from that data, and the authors state that it can be fine-tuned to new sensors and action spaces in a few hours on consumer GPUs [Ref. 4]Sourced.

3. What is inside a VLA — from image, words and state to action

Let us look at the basic shape in a figure. There are three inputs: an image, an instruction in words and the robot’s state (joint angles and so on). A vision-language model (VLM) pre-trained on web images and text receives them, and an action output part at the end produces several steps of actions.

Original SVG, detailed explanation and sources

Original SVG reference: Conceptual diagram (vector drawing). A model that takes images, words and the robot’s state and returns several steps of actions. Building on knowledge learned from web images and text is what VLAs since RT-2 [Ref. 2] have in common. The figure is a rough arrangement, not the internal structure of any specific model.

(1) RT-2: treating actions as “a kind of language”

RT-2 co-fine-tunes a web-scale VLM on robot trajectory data and on web vision-language tasks (such as question answering about images) and has it output robot actions as text tokens, in the same format as language. The paper calls such models VLAs. It reports 6,000 evaluation trials, with generalisation to new objects and the interpretation of commands not in the training data [Ref. 2]Sourced.

(2) OpenVLA: an open 7B VLA

OpenVLA is an open VLA with 7 billion (7B) parameters, trained on 970,000 real-robot demonstrations. It combines the Llama 2 language model with visual features from DINOv2 and SigLIP. It reports 16.5% higher absolute success than the 55B RT-2-X across 29 tasks (with one-seventh of the parameters), and 20.4% higher than Diffusion Policy when fine-tuned to new settings (values under the paper’s settings) [Ref. 5]Sourced.

(3) π0 and GR00T N1: a dedicated part for actions

• π0 adds an action expert of about 300 million parameters to the 3B VLM (PaliGemma), for 3.3 billion parameters in all. It generates actions as continuous values by flow matching, and handles chunks of 50 steps of actions (action chunks) at up to 50 Hz [Ref. 7]Sourced
• GR00T N1 has two systems. The VLM, System 2, interprets the environment and the instruction (10 Hz on an L40 GPU), and a diffusion transformer, System 1, generates smooth motion at a high frequency (120 Hz). The released GR00T-N1-2B has about 2.2 billion parameters. It reports training on a mixture of real-robot trajectories, human video and synthetic data [Ref. 11]Sourced

4. Three ways of outputting actions, and a word on speed

(1) Emit as tokens: cut into 256 bins

RT-2 cuts each action dimension (except the termination command) into 256 bins and expresses one action as 8 integers [Ref. 2]Sourced. OpenVLA also cuts each dimension into 256 bins, dividing evenly the range between the 1st and 99th percentiles of the training data so that extreme outliers do not make the bins coarse [Ref. 5]Sourced.

(2) Emit compressed: FAST

Simple binning was found to produce too many tokens, and to learn poorly, for high-frequency data such as 50 Hz. FAST converts the action sequence into frequency components with a discrete cosine transform, compresses it, and expresses it with few tokens [Ref. 8]Sourced.

As the paper’s table shows, the number of tokens under simple binning equals dimensions x control frequency (35 for 7 dimensions at 5 Hz, 700 for 14 dimensions at 50 Hz)Our calculation. If one second of actions is written out as tokens, ten times the frequency means ten times the tokens. For FAST, the paper reports that under the same conditions 700 tokens for 14 dimensions at 50 Hz become 53, a difference of about 13-fold [Ref. 8]Sourced.

(3) Generate continuous values directly: flow matching and diffusion

The flow matching of π0 and the diffusion transformer of GR00T N1 are generation methods that start from noise and gradually extract a smooth action sequence [Ref. 7, 11]. Since they do not convert to tokens, they can output a whole chunk of actions at once, even at high frequency. π0 reports an on-board inference time of 73 ms (image processing, observation processing, ten iterations to generate actions, and network latency together) [Ref. 7]Sourced.

Putting speed into numbers

To run at a high frequency, the model’s “decision interval” matters. From the values in the papers we calculate the interval of one decision (Our calculation).

Our calculation. Decision interval = 1 / frequency. The computers, robots and tasks differ, so the numbers cannot be ranked. We use the values stated in the papers as they are.

What this shows is that a method that outputs chunks of actions can run a robot at a high frequency even if the model is slow. For example, even with two inferences per second, outputting 25 steps at a time makes a 50 Hz motion possible. The price is that, in the middle of a chunk, the robot cannot react to new observations (commentary).

5. Reading the main models — what they report and how they are released

The table below lines up the main models in terms of what each paper reports and how it is released. Because the evaluation conditions differ, the numbers in different rows cannot be compared to rank the models.

The arrangement of the table is this article’s own (commentary); all entries are Sourced from the papers and official material. Release status was checked on 25 October 2026 within what the papers state and what GitHub shows publicly. “Not checked here” does not mean “does not exist”.

6. Reading benchmarks — LIBERO as an example

(1) What LIBERO is

LIBERO is a simulated benchmark for studying “lifelong learning” in robot manipulation. It provides four task suites (130 tasks in total) and demonstrations teleoperated by people. Spatial, Object and Goal have 10 tasks each, and LIBERO-100 has 100 tasks [Ref. 15]Sourced. VLA papers often use four suites (Spatial, Object, Goal and Long), fine-tune on each suite and report success rates. In the OpenVLA-OFT paper each suite has 10 tasks and 500 expert demonstrations, and evaluation is 500 trials per suite (10 tasks x 50 episodes) [Ref. 6]Sourced.

(2) The “set-up” behind the numbers

• Inputs differ: rows that use only a third-person image sit in the same table as rows that add a wrist-camera image and the robot’s state [Ref. 6]
• Values quoted from other papers: many baselines in the table are values from their original papers, with different implementers and set-ups [Ref. 6]
• Handling of training data: even for the same OpenVLA-OFT, the same paper’s tables give an average of 97.1% with a modified dataset (unsuccessful demonstrations and similar removed) and 94.5% with the original data [Ref. 6]Sourced
• Best result reported: the paper says that checkpoints were tested every 50,000 steps and the best performance of each run was reported [Ref. 6]Sourced
• Fine-tuned per suite: these are not the scores of a general-purpose model solving the tasks “as it is”, but scores after adaptation to each suite [Ref. 6]

(3) Calculating the width of the numbers

With 500 trials per suite, a success rate of 97.1% has a 95% confidence interval (Wilson’s method) of about 95.4-98.3%, and a rate of 95.3% has about 93.0-96.8%Our calculation. This assumes independent trials under identical conditions and expresses only the statistical width. If the difference between rows is smaller than this width, you cannot say which is higher. And rows with different set-ups are not on the same yardstick to begin with.

(4) What was found after “nearly full marks”

Success rates in the 90s have become common on LIBERO, and “saturation” has been pointed out. In October 2025 two reports (preprints) re-evaluated models with conditions changed a little.

• LIBERO-PRO reports that when objects, initial states, instructions and environments are varied within reasonable bounds, the success rate of models above 90% on standard LIBERO falls to 0.0%, and points to a reliance on memorisation of layouts and motion sequences seen in training [Ref. 16]
• LIBERO-Plus reports that small changes of camera viewpoint or the robot’s initial state can drop performance from 95% to below 30%, and states that models barely react to changes in the language instruction and tend to ignore the instruction [Ref. 17]

Both are preprints, and we did not check their peer-review status. The models they evaluated may also differ from the one you use. What can be said is the general caution that a high success rate on a benchmark is a result under that benchmark’s set-up and need not hold when conditions change (commentary).

(5) Seven questions for reading a success rate

Other benchmarks are read with the same questions. CALVIN is a simulated benchmark of language-conditioned, long-horizon manipulation [Ref. 18]. SIMPLER consists of simulation environments paired with real-robot policy evaluation, and reports a strong correlation with real-robot results for particular robot set-ups [Ref. 19]. RoboArena is a distributed real-robot evaluation in which evaluators freely choose tasks and environments and compare two policies double-blind; it ranked seven policies through more than 600 pairwise comparisons [Ref. 20]Sourced.

7. How to view demonstration videos and the word “general-purpose”

Videos are often used to introduce robots. A video can be shot by choosing the scene to show, so we do not treat it as evidence of capability. Check the following (this is this article’s arrangement).

• Autonomous or teleoperated: is the robot moving by itself under a policy, or is a person operating it remotely? Is the motion a fixed sequence practised in advance? See whether the video’s description says “autonomous”, “1x speed” or “no teleoperation”
• One run out of how many: has a successful run been cut out? Are the success rate and number of trials given in a paper or technical document?
• Editing and speed: is there any fast-forwarding or cutting and splicing?
• Is the environment new: is it the same place and objects used in training? π0.5 describes evaluation in “new homes”, but the conditions (whether people intervened, for example) must be checked in the paper [Ref. 9]

How demonstration data are collected (teleoperation), and how to tell autonomous from teleoperated runs, are covered in detail in our article on teleoperation and data collection. Note that the training data of GR00T N1 include teleoperated data collected by people operating devices [Ref. 11]. Training data being teleoperated is a separate matter from a robot at evaluation time running autonomously, but it is worth checking whether the material says which one it is.

8. A materials engineer’s view 2: a difference in “body” is a difference in materials and parts

9. Limits, and what could not be confirmed / summary

(1) Comparing model performance

Success rates and processing speeds are values reported for the robots, tasks, environments and computers each paper chose. We did not confirm independent third-party reproduction. We do not rank models.

(2) The inside of undisclosed models

For models whose weights are not released, such as Gemini Robotics, internal details can be confirmed only within the technical reports. We limited release status to what the papers state and what GitHub shows publicly, and did not check the availability of companies’ services.

(3) Preprints

LIBERO-PRO and LIBERO-Plus are preprints from October 2025, and we did not check their peer-review status. Their results are reports on the models those papers evaluated.

(4) The text of standards

For ISO standards we confirmed only number, title and stage and did not read the text. We could not confirm a published standard that directly defines how to assess the safety of learned policies (VLAs).

(5) What we did not cover

World models are in our article on world models and embodied AI, and humanoid products and businesses in our article on humanoid robots. We do not cover market size, prices or share prices.

10. How to read the yardsticks and standards

A VLA is basically evaluated by the success rate on tasks, measured on simulated benchmarks and on real robots. In practice, robustness when conditions change, speed of inference (latency) and the amount of data needed for fine-tuning also act as yardsticks. Common standards are still few, and safety rests on existing standards such as those for industrial robots (checked 25 October 2026).

The arrangement of the table is Our calculation (organising). Numbers, titles and stages of ISO standards were checked on the official pages (25 October 2026). The text of the standards is paid, so we stay with titles and stages.

Notes on reading (for materials engineers)

• Demonstrations, benchmarks and the shop floor are different things: videos and benchmarks give values for chosen tasks in tidy environments. Re-measure with your own objects, lighting, floor and a worn machine
• A developer’s report and independent verification are different: most VLA results are reported by the group or company that developed the model. Check whether independent evaluation (such as RoboArena) exists
• Is a standard a requirement, information, or in preparation? ISO 10218 sets safety requirements, ISO/IEC TR 5469 provides information, and ISO/IEC TS 22440 is in preparation. “Conforms to a standard” does not mean a learned policy is safe
• Plans and schedules are not settled: companies’ “coming soon” or “mass production” statements are Not yet confirmed. We do not cover them here

11. Glossary

12. References

1. Brohan A. et al. “RT-1: Robotics Transformer for Real-World Control at Scale”, arXiv:2212.06817 (2022) — arxiv.org/abs/2212.06817
2. Brohan A. et al. “RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control”, arXiv:2307.15818 (2023) — arxiv.org/abs/2307.15818
3. Open X-Embodiment Collaboration “Open X-Embodiment: Robotic Learning Datasets and RT-X Models”, arXiv:2310.08864 (2023) — arxiv.org/abs/2310.08864
4. Octo Model Team “Octo: An Open-Source Generalist Robot Policy”, arXiv:2405.12213 (2024) — arxiv.org/abs/2405.12213
5. Kim M. J. et al. “OpenVLA: An Open-Source Vision-Language-Action Model”, arXiv:2406.09246 (2024) — arxiv.org/abs/2406.09246
6. Kim M. J., Finn C., Liang P. “Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success” (OpenVLA-OFT), arXiv:2502.19645 (2025) — arxiv.org/abs/2502.19645
7. Black K. et al. (Physical Intelligence) “π0: A Vision-Language-Action Flow Model for General Robot Control”, arXiv:2410.24164 (2024) — arxiv.org/abs/2410.24164
8. Pertsch K. et al. “FAST: Efficient Action Tokenization for Vision-Language-Action Models”, arXiv:2501.09747 (2025) — arxiv.org/abs/2501.09747
9. Physical Intelligence et al. “π0.5: a Vision-Language-Action Model with Open-World Generalization”, arXiv:2504.16054 (2025) — arxiv.org/abs/2504.16054
10. Physical Intelligence GitHub repository “openpi” (Apache License 2.0). Repository and licence confirmed through GitHub’s public API (25 October 2026); the contents of the released weights were not checked here — github.com/Physical-Intelligence/openpi
11. NVIDIA et al. “GR00T N1: An Open Foundation Model for Generalist Humanoid Robots”, arXiv:2503.14734 (2025) — arxiv.org/abs/2503.14734
12. Gemini Robotics Team “Gemini Robotics: Bringing AI into the Physical World”, arXiv:2503.20020 (2025) — arxiv.org/abs/2503.20020
13. Gemini Robotics Team “Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion Transfer”, arXiv:2510.03342 (2025) — arxiv.org/abs/2510.03342
14. Shukor M. et al. “SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics”, arXiv:2506.01844 (2025) — arxiv.org/abs/2506.01844
15. Liu B. et al. “LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning”, arXiv:2306.03310 (2023; NeurIPS 2023 Datasets and Benchmarks) — arxiv.org/abs/2306.03310
16. Zhou X. et al. “LIBERO-PRO: Towards Robust and Fair Evaluation of Vision-Language-Action Models Beyond Memorization”, arXiv:2510.03827 (2025; preprint). A preprint; its peer-review status was not checked here — arxiv.org/abs/2510.03827
17. Fei S. et al. “LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models”, arXiv:2510.13626 (2025; preprint). A preprint; its peer-review status was not checked here — arxiv.org/abs/2510.13626
18. Mees O. et al. “CALVIN: A Benchmark for Language-Conditioned Policy Learning for Long-Horizon Robot Manipulation Tasks”, arXiv:2112.03227 (2021) — arxiv.org/abs/2112.03227
19. Li X. et al. “Evaluating Real-World Robot Manipulation Policies in Simulation” (SIMPLER), arXiv:2405.05941 (2024) — arxiv.org/abs/2405.05941
20. Atreya P. et al. “RoboArena: Distributed Real-World Evaluation of Generalist Robot Policies”, arXiv:2506.18123 (2025) — arxiv.org/abs/2506.18123
21. ISO ISO 10218-1:2025 “Robotics — Safety requirements — Part 1: Industrial robots” (Edition 3, February 2025). Checked 25 October 2026 — iso.org/standard/73933.html
22. ISO ISO 10218-2:2025 “Robotics — Safety requirements — Part 2: Industrial robot applications and robot cells” (Edition 2, February 2025). Checked 25 October 2026 — iso.org/standard/73934.html
23. ISO/IEC ISO/IEC TR 5469:2024 “Artificial intelligence — Functional safety and AI systems” (Edition 1, January 2024). Checked 25 October 2026 — iso.org/standard/81283.html
24. ISO/IEC ISO/IEC CD TS 22440-1 “Artificial intelligence — Functional safety and AI systems — Part 1: Requirements” (committee draft; stage 30.60). Checked 25 October 2026 — iso.org/standard/89535.html

13. Claim-to-source audit

Last updated: 25 October 2026 / Sources are limited to primary material (arXiv papers, public GitHub information and the official ISO pages). We could not read the text of the ISO standards because it is paid. LIBERO-PRO and LIBERO-Plus are preprints. The numerical examples are arithmetic on values from the papers and are not the performance of any real machine. We could not confirm, and so do not state, independent third-party reproduction of each model, rankings among models, the internal structure of undisclosed models, or a standard that directly defines the safety assessment of learned policies. Figures 1 and 8 are ImageGen-generated conceptual illustrations; the other figures are explanatory concept diagrams or figures containing our calculations.