1. What teleoperation and data collection are (the short version)

• Teleoperation has two faces: (1) people doing work from a distance, in dangerous or far-away places (real work); (2) a means of recording examples to teach robots (data collection). This article is mainly about (2)
• Variety can matter more than quantity: from over 40,000 demonstrations and over 15,000 real rollouts, Lin et al. report that generalisation improved roughly as a power law of the number of environments and objects, and that the variety of environments and objects mattered more than the total number of demonstrationsSourced[Ref. 18]
• A video alone does not show whether it is autonomous or teleoperated: companies’ official pages include one that states “uncut, from a single policy operating fully autonomously”, one that says it worked closely with operators during data collection, and one whose text did not let us confirm eitherSourced[Ref. 22, 23, 24]. What counts is the stated conditions

2. What teleoperation is — the border with autonomy

Teleoperation is a person operating a robot from a distance. The person’s hand motion becomes the robot’s motion directly, so the robot is “an extension of the person’s limbs”. Autonomy, by contrast, is the robot moving on its own judgement (a policy). Between them there are several stages.

(1) Returning the operator’s “feel”: bilateral teleoperation

A scheme in which the force the robot feels on touching something is returned to the operator’s hand is called bilateral teleoperation. A historical survey is the paper by Hokayem and Spong [Ref. 1] (we confirmed only title and bibliographic data, not the contents). When images and force are returned to the operator, fine work is thought to become easier, but it becomes hard when the delay is large (commentary).

(2) Delay in numbers

The farther away the operator, the longer a signal takes to go and come back. Assuming the speed of light in optical fibre to be about 200,000 km/s, we calculate the round-trip delay from propagation alone (Our calculation).

Our calculation. Round-trip propagation time = 2 x distance / speed of light. The speed in optical fibre is assumed to be about 200,000 km/s (about two-thirds of that in vacuum). Route curvature, equipment processing and video compression times are not included.

How delay matters depends greatly on the task and the system. RoboTurk is a system that collects data from remote workers using smartphones, and it reports that poor network conditions such as low bandwidth and high delay did not substantially affect the remote workers’ ability to carry out demonstrations successfully [Ref. 2]Sourced. That is a result for the paper’s tasks and set-up and need not hold for precise work with force feedback (commentary).

3. Four ways to collect demonstrations

Ways of collecting demonstrations can be sorted by “whether a robot is used” and “what is measured of the person”.

Original SVG, detailed explanation and sources

Original SVG reference: Conceptual diagram (vector drawing). A: GELLO is a device with the same kinematic structure as the target robot arm, built from 3D-printed parts and off-the-shelf motors [Ref. 5]. B: [Ref. 6, 7]. C: [Ref. 9]. D: [Ref. 10, 11, 12].

A. Leader-follower (a robot plus a device of the same structure)

The robot (follower) reproduces the motion of a “leader” device moved by a person. ACT (ALOHA) collected demonstrations with low-cost hardware and its own teleoperation device and reports learning six difficult tasks at 80-90% success with 10 minutes of demonstrations [Ref. 3]. GELLO is a way of building a device with the same kinematic structure as the target robot arm from 3D-printed parts and off-the-shelf motors, and a user study reports that it collected demonstrations more reliably and efficiently than VR controllers and 3D spacemouses [Ref. 5]Sourced.

B. VR and hand tracking

Open-TeleVision is an immersive teleoperation system in which the operator can look around the robot’s surroundings actively in stereo. It collected data on four long tasks (can sorting, insertion, folding and unloading) with two kinds of humanoid and deployed the learned policies on the real machines [Ref. 6]. AnyTeleop handles vision-based teleoperation of dexterous arm-hand systems in one system across several robots and camera set-ups [Ref. 7]. HumanPlus tracks a person’s whole-body and hand motion with only an RGB camera so that a humanoid “shadows” it like a shadow; it collected whole-body data this way and reports that, with up to 40 demonstrations, the humanoid carried out tasks autonomously at 60-100% success [Ref. 8]Sourced.

C. Gloves and motion capture

DexCap is a portable hand motion capture system based on SLAM and electromagnetic fields. It tracks wrist and finger motion resistant to occlusion and also records the 3D surroundings. It also proposed a method of learning dexterous robot-hand policies from human hand-motion data (DexIL) and a mechanism for a person to correct the policy during execution, and was evaluated on six dexterous tasks [Ref. 9]Sourced.

D. Hand-held tools and human video

UMI uses hand-held grippers to collect in-the-wild demonstrations without a robot and transfers them directly to robot policies. With inference-time latency matching and a relative-trajectory action representation, it says the policies become hardware-agnostic [Ref. 10]. EgoMimic uses egocentric video of people taken with Project Aria glasses together with 3D hand tracking, and co-trains on human and robot data as the same kind of “demonstration data”. It reports a scaling trend in which adding one hour of hand data is significantly more valuable than adding one hour of robot data [Ref. 11]. Ego4D is a dataset of 3,670 hours of everyday egocentric video from 931 camera wearers, and is not robot demonstrations as such [Ref. 12]Sourced.

4. Comparing the ways — what is gained and what is given up

The table is this article’s arrangement (commentary), not a recommendation by the papers. We do not say which method is better. In practice many examples combine several methods.

5. Quantity and quality of data — how many items, how many hours, how varied

(1) The scale of main datasets

All Sourced (abstracts of the papers). Robots, tasks and collection methods differ, so the numbers cannot be lined up to rank them.

(2) Counting the human time

Demonstrations are made out of human time. If we assume that one demonstration takes 40 s of operation plus 20 s of resetting, 60 s in all, then 100 demonstrations take about 1.7 hours, 1,000 about 16.7 hours, 10,000 about 167 hours, 100,000 about 1,667 hours and 1 million about 16,667 hours (Our calculation). DROID has 76,000 trajectories in 350 hours, which is about 16.6 s of “interaction time” per trajectory (350 hours / 76,000) [Ref. 13].

The GR00T N1 paper says that robot data scale linearly with human labour because a human operator has to teleoperate the robot to produce each trajectory, and reports expanding its 88 hours of in-house teleoperated data about tenfold to 827 hours using video-generation models [Ref. 21]Sourced. Ways of reducing human time (video, synthesis, using human video) are a major concern in data collection.

(3) “Variety” rather than “quantity”

Even for the same 1,000 demonstrations, 1,000 in one place and 100 in each of ten places have different content. Lin et al. report that generalisation improved roughly as a power law of the number of environments and objects, that the variety of environments and objects mattered more than the total number of demonstrations, and that once the number of demonstrations per environment or object passed a threshold, extra demonstrations had little effect. They also state that data collected by four people in one afternoon brought two tasks to about 90% success in new environments with unseen objects [Ref. 18]Sourced.

(4) Data quality and records

The quality of demonstration data includes the operator’s skill, habits, fatigue and the variation of conditions on the day. The ACT paper names non-stationary human demonstrations as a challenge [Ref. 3], and robomimic analyses dependence on demonstration quality [Ref. 19]. So if you record who, when, on which robot, in which environment and with which objects, success or failure, and whether a person intervened, you can later trace the causes of failure (commentary). As a shared data format, Open X-Embodiment pooled data from 21 institutions in a standard format [Ref. 17]. An open framework for learning is LeRobot (Apache-2.0) [Ref. 25].

6. How to tell whether a demonstration is autonomous or teleoperated (important)

When you come across a robot video, it is good practice to ask “is this autonomous?” But do not judge by appearance. Smooth motion can be teleoperated, and clumsy motion can be autonomous. What you check is the stated conditions (this is this article’s arrangement).

The table is this article’s arrangement (commentary). Absence of a statement does not mean the run is teleoperated; “could not be confirmed” is the appropriate wording.

Examples of how companies’ official pages word it (checked 25 October 2026)

Here are examples of how actual official pages describe autonomy, teleoperation and speed. They do not judge the companies, nor whether a given video is actually autonomous.

• Physical Intelligence (31 October 2024): for its laundry video it states “uncut, from a single policy operating fully autonomously”. Whether the evaluation videos are autonomous or teleoperated, and the playback speed, could not be confirmed from the text [Ref. 22]Sourced
• Figure (26 February 2025): for logistics work it describes “working closely with the teleoperators to refine and uniformise manipulation strategies” to train the policy. It also says a test-time method of speeding up execution can run faster than the demonstrators (up to 50% speed-up). In the text we could not confirm whether each video is autonomous or its playback speed [Ref. 23]Sourced
• Google DeepMind (12 March 2025): in the text of the blog, the video descriptions are brief captions, and we could not confirm whether the videos are autonomous or teleoperated, or the playback speed [Ref. 24]Sourced. Detailed conditions have to be checked in the technical report

These differences in wording also reflect different purposes of the blog posts. An introductory blog does not necessarily state every condition. That is why the honest way to read is to check papers, technical reports and specifications, and to write “could not be confirmed” where nothing is stated (commentary).

The counting rule changes the “success rate”

When people intervened, the way of counting changes the number. As an assumption, suppose that of 100 task attempts, 70 were completed autonomously, 18 were completed after a person intervened and 12 failed.

• Counting every completion, intervention or not, as a success gives 88% (70 + 18)
• Counting only completions without human help gives 70%
• The share of attempts in which a person intervened is 18% (18 / 100)

Our calculation. The breakdown is an illustrative assumption, not the result of any real robot.

7. A materials engineer’s view: quality of data, gloves and the materials of devices

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

(1) Comparing results in papers

Dataset sizes, success rates and collection times are values reported under the conditions each paper chose. We did not confirm independent third-party reproduction. We do not rank them.

(2) Companies’ videos

The wording of companies’ official pages is a set of examples within the text we checked, and “could not be confirmed” does not mean a run is teleoperated. We did not check the autonomy of all videos and announcements of each company. We did not include cases, such as 1X, for which we could not reach the official page.

(3) The text of standards

For ISO standards we confirmed only number, title and publication date and did not read the text. We could not confirm an international standard that defines classes of teleoperation and autonomy. We did not open the ASTM site. The NIST “Response Robots” page says it is developing test methods and performance metrics for remotely operated robots (including operator proficiency), but it does not give numbers or titles of individual standards.

(4) What we did not cover

Remote operation of medical robots (surgical assistance) is in our article on surgical robots, and remote operation in space and underwater in the respective articles. We do not cover market size or prices.

9. How to read the yardsticks and standards

The yardsticks for data collection are the quantity, variety and quality of data and the success rate of the policy learned from them. For teleoperation itself, there are test methods, including operator proficiency, for rescue-type robots. We could not confirm a common classification that defines autonomous versus teleoperated (checked 25 October 2026).

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

Notes on reading (for materials engineers)

• Demonstrations, benchmarks and the shop floor are different things: success rates on videos and datasets are values for chosen tasks and environments. Re-measure with your own objects and environment
• A developer’s statement and independent verification are different: whether a video is autonomous and its success rate are the developer’s wording. Check whether independent evaluation exists
• Is a standard a requirement or information? ISO/IEC 5259 is a framework for data quality and ISO 10218 sets safety requirements for industrial robots; neither directly defines the quality of robot demonstration data or whether a video is autonomous
• Plans and schedules are not settled: companies’ “coming soon” or “mass production” statements are Not yet confirmed. We do not cover them here

10. Glossary

11. References

1. Hokayem P. F., Spong M. W. “Bilateral teleoperation: An historical survey”, Automatica 42(12), 2035–2057 (2006), DOI 10.1016/j.automatica.2006.06.027. Title and bibliographic data checked in Crossref; contents not checked here — doi.org/10.1016/j.automatica.2006.06.027
2. Mandlekar A. et al. “RoboTurk: A Crowdsourcing Platform for Robotic Skill Learning through Imitation”, arXiv:1811.02790 (2018) — arxiv.org/abs/1811.02790
3. Zhao T. Z. et al. “Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware” (ALOHA, ACT), arXiv:2304.13705 (2023) — arxiv.org/abs/2304.13705
4. Fu Z., Zhao T. Z., Finn C. “Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation”, arXiv:2401.02117 (2024) — arxiv.org/abs/2401.02117
5. Wu P. et al. “GELLO: A General, Low-Cost, and Intuitive Teleoperation Framework for Robot Manipulators”, arXiv:2309.13037 (2023) — arxiv.org/abs/2309.13037
6. Cheng X. et al. “Open-TeleVision: Teleoperation with Immersive Active Visual Feedback”, arXiv:2407.01512 (2024) — arxiv.org/abs/2407.01512
7. Qin Y. et al. “AnyTeleop: A General Vision-Based Dexterous Robot Arm-Hand Teleoperation System”, arXiv:2307.04577 (2023) — arxiv.org/abs/2307.04577
8. Fu Z. et al. “HumanPlus: Humanoid Shadowing and Imitation from Humans”, arXiv:2406.10454 (2024) — arxiv.org/abs/2406.10454
9. Wang C. et al. “DexCap: Scalable and Portable Mocap Data Collection System for Dexterous Manipulation”, arXiv:2403.07788 (2024) — arxiv.org/abs/2403.07788
10. Chi C. et al. “Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots”, arXiv:2402.10329 (2024) — arxiv.org/abs/2402.10329
11. Kareer S. et al. “EgoMimic: Scaling Imitation Learning via Egocentric Video”, arXiv:2410.24221 (2024) — arxiv.org/abs/2410.24221
12. Grauman K. et al. “Ego4D: Around the World in 3,000 Hours of Egocentric Video”, arXiv:2110.07058 (2021) — arxiv.org/abs/2110.07058
13. Khazatsky A. et al. “DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset”, arXiv:2403.12945 (2024) — arxiv.org/abs/2403.12945
14. Walke H. et al. “BridgeData V2: A Dataset for Robot Learning at Scale”, arXiv:2308.12952 (2023) — arxiv.org/abs/2308.12952
15. Fang H.-S. et al. “RH20T: A Comprehensive Robotic Dataset for Learning Diverse Skills in One-Shot”, arXiv:2307.00595 (2023) — arxiv.org/abs/2307.00595
16. AgiBot World Contributors “AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Systems”, arXiv:2503.06669 (2025) — arxiv.org/abs/2503.06669
17. Open X-Embodiment Collaboration “Open X-Embodiment: Robotic Learning Datasets and RT-X Models”, arXiv:2310.08864 (2023) — arxiv.org/abs/2310.08864
18. Lin F. et al. “Data Scaling Laws in Imitation Learning for Robotic Manipulation”, arXiv:2410.18647 (ICLR 2025) — arxiv.org/abs/2410.18647
19. Mandlekar A. et al. “What Matters in Learning from Offline Human Demonstrations for Robot Manipulation”, arXiv:2108.03298 (2021) — arxiv.org/abs/2108.03298
20. Luo J. et al. “Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning” (HIL-SERL), arXiv:2410.21845 (2024) — arxiv.org/abs/2410.21845
21. NVIDIA et al. “GR00T N1: An Open Foundation Model for Generalist Humanoid Robots”, arXiv:2503.14734 (2025) — arxiv.org/abs/2503.14734
22. Physical Intelligence Blog post “π0: Our First Generalist Policy” (31 October 2024). Text checked on 25 October 2026 — pi.website/blog/pi0
23. Figure AI News post “Helix Accelerating Real-World Logistics” (26 February 2025). Text checked on 25 October 2026 — figure.ai/news/helix-logistics
24. Google DeepMind Blog post “Gemini Robotics brings AI into the physical world” (12 March 2025). Text checked on 25 October 2026 — deepmind.google/discover/blog/gemini-robotics-brings-ai-into-the-physical-world
25. Hugging Face GitHub repository “huggingface/lerobot” (Apache License 2.0). Licence confirmed through GitHub’s public API (25 October 2026) — github.com/huggingface/lerobot
26. NIST “Response Robots” (standard test methods and performance metrics for remotely operated robots). Checked 25 October 2026 — nist.gov/el/intelligent-systems-division-73500/response-robots
27. ISO/IEC ISO/IEC 5259-1:2024 “Artificial intelligence — Data quality for analytics and machine learning (ML) — Part 1: Overview, terminology, and examples” (July 2024). Checked 25 October 2026 — iso.org/standard/81088.html
28. ISO/IEC ISO/IEC 5259-3:2024 “Artificial intelligence — Data quality for analytics and machine learning (ML) — Part 3: Data quality management requirements and guidelines” (July 2024). Checked 25 October 2026 — iso.org/standard/81092.html
29. 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

12. Claim-to-source audit

Last updated: 25 October 2026 / Sources are limited to primary material (arXiv papers, companies’ official pages, the official NIST and ISO pages and public GitHub information). We could not read the text of the ISO standards because it is paid. We did not open the ASTM site. The numerical examples are calculations from values we assumed and are not the results of any real robot or operator. We could not confirm, and so do not state, an international standard that defines classes of teleoperation and autonomy, whether each company video is autonomous or teleoperated, or statistics on human intervention. Figures 1 and 8 are ImageGen-generated conceptual illustrations; the other figures are explanatory concept diagrams or figures containing our calculations.