# Teleoperation

> Source: https://aiwiki.ai/wiki/teleoperation
> Updated: 2026-07-24
> Fact-checked: 2026-07-24
> Categories: AI History, Embodied AI, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Teleoperation." aiwiki.ai, 24 Jul 2026. https://aiwiki.ai/wiki/teleoperation
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

Teleoperation is the operation of a machine by a human at a distance, with commands carried over a communication link and the machine's cameras, microphones and force sensors standing in for the operator's own senses. It covers a nuclear technician working through a shielded wall, a surgeon at a console a few meters from the patient, a rover on the Moon driven from a control room in Crimea, and a robot arm in a laboratory puppeteered from a phone on another continent. What separates teleoperation from simple remote control is the closed loop: the operator keeps perceiving the remote scene and correcting, instead of issuing a command and waiting to see what happened.

The field grew out of the need to handle radioactive material after the Second World War, and its central engineering problems have changed very little since. Delay in the communication link destabilizes force feedback [5]. Cameras give a narrow, flattened view of a workspace the operator cannot walk around. Contact forces, which people use constantly without noticing, are expensive to measure and awkward to transmit. Most of the theory in the field exists to manage those three deficits.

Since about 2023 teleoperation has acquired a second identity inside artificial intelligence. It is now a primary source of the training data behind robot manipulation policies: [imitation learning](https://aiwiki.ai/wiki/imitation_learning) needs demonstrations expressed in the robot's own action space, and a human puppeteering the robot is the standard way to produce them [25][33][36]. The same machinery doubles as a commercial fallback, because when a [humanoid robot](https://aiwiki.ai/wiki/humanoid_robot) or a robotaxi hits a situation its policy cannot resolve, a remote human either takes over or advises. Both uses put teleoperation in the middle of arguments about how autonomous the current generation of "autonomous" systems really is.

## Origins in nuclear remote handling

The modern master and slave manipulator was designed by Raymond C. Goertz, a mechanical engineer who joined Argonne National Laboratory in 1947 after seven years at Sperry Gyroscope's Servomechanisms Laboratory [1]. Argonne needed a way to handle irradiated material inside shielded hot cells, where an operator could look through leaded glass but could not reach in. Goertz's answer, designed in 1948, was a seven degree of freedom bilateral pantograph device that transmitted motion through metal tapes; he filed the patent, "Remote-control manipulator," on 16 December 1949, and it was granted as U.S. Patent 2,632,574 on 24 March 1953 [1][2]. A 1951 revision coupled the master and slave arms through steel pulleys and cables [1].

The mechanical linkage limited how far apart the two arms could be, so Goertz went electrical. Applying the principles of cybernetics, he built the first electrical master and slave manipulator system, which he and W. M. Thompson described in the November 1954 issue of Nucleonics [1][46]. By 1954 a modified version of the master and slave manipulator, the CRL Model 8, had begun commercial production [1]. Goertz had also concluded that a haptic sense was necessary for handling delicate objects and used force feedback to improve the deftness of the human and machine combination: in his formulation a bilateral coupling means that forces at the slave end are reflected at the master end, and that displacements at the slave end can produce displacements at the master [1]. That insistence on reflecting force back to the hand is the direct ancestor of essentially every bilateral teleoperator built since.

The vocabulary has shifted. Contemporary systems, including the robot learning rigs described below, generally label the operator side the leader and the remote side the follower [28].

## Bilateral control and the delay problem

A unilateral teleoperator sends position commands one way. A bilateral one closes a second loop, reflecting the forces the follower encounters back into the operator's hand. Bilateral control is what makes fine contact work possible, and it is also what makes delay dangerous, because a force signal that arrives late can pump energy into the loop instead of dissipating it.

William Ferrell measured the human consequence in 1965. He introduced transmission delay between the master and slave elements of a manipulator and measured the effect on performance for both simple and complex tasks; operators adopted a "move and wait" strategy to cope with it, making a small motion, stopping, waiting for the display to catch up, evaluating, then moving again [3]. The strategy is safe and extremely slow, and it still describes how people behave on high latency links today.

Two lines of work attacked the problem. The first was to change the control law. R. J. Anderson and M. W. Spong showed in 1989 that passivity and scattering theory explain why conventional bilateral controllers go unstable for certain environments, and derived a control law that stays stable for any environment despite the transmission delay [5]. Gunter Niemeyer and Jean-Jacques Slotine extended the idea in 1991 with wave variables, a passivity based encoding of the transmitted signals that conserves energy across a delayed channel and yields a workable force reflecting configuration [6]. Wave variable methods remain standard in delayed bilateral control.

The second line was to stop showing the operator the delayed truth and show a prediction instead. Thomas Sheridan's framing of human supervisory control, set out in his 1992 book Telerobotics, Automation, and Human Supervisory Control, treats the human as intermittently commanding and monitoring a partly autonomous machine rather than continuously driving it [4]. Predictive displays render a simulated model of the remote scene in real time, letting the operator act against the model while the real robot catches up.

Force display hardware developed alongside. Thomas Massie and J. K. Salisbury's PHANToM, presented at the ASME Winter Annual Meeting in 1994, measured a user's fingertip position and exerted a controlled force vector back on it through three small motors, with the user's index finger in a thimble [7]. It made desktop haptic interaction cheap enough to become a research commodity.

The delays that real systems have had to absorb range from tens of milliseconds to several seconds.

| System | Year | Round-trip delay handled | Method |
| --- | --- | --- | --- |
| ROTEX, Spacelab D-2 to ground [10] | 1993 | 5-7 seconds | Predictive computer graphics simulation |
| Lindbergh Operation, New York to Strasbourg [19] | 2001 | About 155 ms, against an assumed safe limit of 330 ms | Dedicated fiber optic circuit |
| Analog-1, ISS to a rover in the Netherlands [14] | 2019 | About 800 ms average, outliers to 3 seconds, with packet loss | Force feedback controller tolerant of variable delay |
| COBALT cloud teleoperation [40] | 2026 | Sub-100 ms end to end for up to 8 concurrent users per GPU | In-memory data cache and efficient video streaming |

## Space telerobotics

The Soviet Lunokhod rovers were the first substantial planetary teleoperation. Lunokhod 1 launched on 10 November 1970 and landed on 17 November; in 322 Earth days of operation it traveled 10,540 meters and returned more than 20,000 television images and 206 high resolution panoramas [8]. Driving was done from a tracking facility at the closed town of Simferopol-28 in Crimea, and Lunokhod 2 was handled by a five-man team of controllers [9]. The constraint was not only signal travel time: Lunokhod 2's cameras returned high resolution images at 3.2, 5.7, 10.9 or 21.1 seconds per frame, so the crew were steering on something closer to a slide show than a video feed [9].

The German ROTEX experiment flew on Space Shuttle Columbia's STS-55 mission, the Spacelab D-2 flight that launched on 26 April 1993 and landed on 6 May [11]. ROTEX exercised four operating modes: preprogrammed automatic execution, on-board teleoperation by an astronaut using stereo television, ground teleoperation using predictive computer graphics, and tele-sensor-programming, described as learning by demonstration in a simulated environment on the ground [10]. The predictive graphics compensated for a communication delay of 5 to 7 seconds, which was enough for controllers in Germany to carry out the first telerobotic capture of a free floating object in space [10][11]. ROTEX also assembled a mechanical grid structure and connected and disconnected an orbital replaceable unit with a bayonet closure [10].

NASA's Robonaut program aimed at something different: a dexterous humanoid torso that could work alongside crew and take on hazardous tasks in their place. Robonaut 1 was built at the Johnson Space Center's Dexterous Robotics Laboratory with DARPA among the partners, and never flew [13]. A 2007 Space Act Agreement brought in General Motors; the resulting Robonaut 2 was revealed in February 2010 and launched to the International Space Station on STS-133 on 24 February 2011, becoming the first humanoid robot in space [12][13]. Both station crew and ground controllers operated it using telepresence, and each arm had seven degrees of freedom and could hold 20 pounds in any pose [12][13]. Climbing legs were added in 2014; NASA announced in April 2018 that R2 would return to Earth for repair [12][13].

The most demanding orbital demonstration of force feedback was ESA's Analog-1. On 25 November 2019 astronaut Luca Parmitano, aboard the ISS, commanded the Interact rover at a test site in Valkenburg in the Netherlands and collected rock samples with its robotic arm [14]. The link carried very high and variable time delays averaging 800 milliseconds with outliers up to 3 seconds, plus packet loss and outright interruptions, and the controller was built so that the robot applies no force to the environment before the astronaut senses it [14].

## Undersea work and nuclear decommissioning

Deep ocean science runs on tethered remotely operated vehicles. Woods Hole Oceanographic Institution's Jason, first launched in 1988 and operated by the National Deep Submergence Laboratory, has a test depth of 6,500 meters and hangs from a 10 kilometer electro-optical-mechanical tether; a companion vehicle, Medea, sits between ship and ROV as a shock absorber, buffering Jason from ship motion while providing lighting and an overhead view [15]. As of 2020 Jason had completed 147 cruises, over 1,200 dives and more than 16,000 hours of dive time [15]. The hybrid vehicle Nereus, whose test depth was 10,902 meters, reached exactly that depth in the Challenger Deep on 31 May 2009 [16]. In its remotely operated configuration it flew on an optical fiber tether about the diameter of a human hair, which could bear only 4 kilograms; the vehicle carried roughly 40 kilometers of that cable and could also be switched to a free swimming autonomous mode [16].

Nuclear decommissioning remains the domain teleoperation was invented for. On 7 November 2024 Tokyo Electric Power Company completed the first trial extraction of melted fuel debris from Unit 2 at Fukushima Daiichi, using a remotely operated telescopic device with a gripper that can extend up to 22 meters through a penetration in the primary containment vessel [17]. The plan was to recover a few grams of debris; the gripper closed on a sample on 30 October, and on 6 November Tepco confirmed the sample's dose rate was less than 24 mSv/h at a distance of 20 centimeters [17]. The operation was not smooth: on 17 September it was discovered that cameras on the end of the telescopic device were not functioning correctly, and they had to be replaced before work could continue [17].

## Surgery

Robotic surgery is teleoperation with the two ends in the same room. Intuitive Surgical's da Vinci system was cleared by the FDA in 2000 for urologic, general laparoscopic, gynecologic laparoscopic and non-cardiovascular thoracoscopic procedures, and for thoracoscopically assisted cardiotomy [18]. The surgeon sits at a console with stereo vision and hand controls and drives a patient-side cart carrying three to four robotic arms, one of which holds the 3D cameras while the others act as scalpels, scissors, electrocautery devices or graspers [18]. In 2012 it was used in an estimated 200,000 surgeries, most often hysterectomies and prostate removals [18].

The one famous long-distance case is the Lindbergh Operation. On 7 September 2001 surgeons working in New York, in a team led by Jacques Marescaux, removed the gallbladder of a 68-year-old woman in Strasbourg using a modified ZEUS system from Computer Motion over a France Telecom fiber optic circuit [19][20][47]. The operation lasted less than an hour across a 14,000 km circuit, and the surgeons' movements appeared on their screens with a delay of around 155 milliseconds, well inside the estimated safe lag of 330 milliseconds; setting up the robot took 16 minutes and the gallbladder dissection itself 54 [19]. The result was published in Nature on 27 September 2001 [20]. Transatlantic telesurgery did not become routine afterwards; the dedicated bandwidth, the liability and the licensing all argued against it.

Force feedback took much longer to reach surgery than it did to reach Goertz's hot cells. Intuitive's da Vinci 5, cleared by the FDA in March 2024, introduced Force Feedback technology, which lets surgeons sense the pressure their instruments apply to tissue; Intuitive says surgeons using it can deliver up to 43 percent less force on tissue, a figure it footnotes to preclinical data on file, and claims 10,000 times the computing power of the da Vinci Xi [21][22]. A further software clearance in September 2025 added an on-screen gauge, working like a speedometer, that displays the force the instruments are applying [21]. That is Goertz's core idea arriving in a different industry seventy years later.

## Driving: remote control versus remote advice

The [autonomous vehicle](https://aiwiki.ai/wiki/autonomous_vehicle) industry draws a sharp line between remote driving and remote assistance, and the distinction matters because the two carry very different safety arguments.

[Waymo](https://aiwiki.ai/wiki/waymo) has been explicit that its fleet response agents do not drive. In a May 2024 post the company described agents viewing live exterior camera feeds and a 3D representation of the vehicle's perception, then answering questions, flagging lane closures or proposing a path for the vehicle to consider [23]. Waymo's position is that "the Waymo Driver does not rely solely on the inputs it receives from the fleet response agent and it is in control of the vehicle at all times," and that it continues acting on its own information while waiting for a reply [23].

True remote driving exists as a commercial product. The German company Vay launched a teledriving service in Las Vegas on 17 January 2024 in which a remote teledriver, seated at a station with a steering wheel, pedals and automotive-grade controls, drives an electric rental car to the customer, who then drives it themselves before a teledriver collects it [24]. Camera feeds go to the operator's screens and microphones pipe road sounds such as sirens into the teledriver's headphones [24].

## Teleoperation as a data source for robot learning

Behavior cloning requires demonstrations in the robot's own action space, which is why teleoperation became the bottleneck resource of modern [robot learning](https://aiwiki.ai/wiki/robot_learning). The clearest demonstration of how far cheap teleoperation could go was [ALOHA](https://aiwiki.ai/wiki/aloha_robot), short for A Low-cost Open-source Hardware System for Bimanual Teleoperation, from Tony Z. Zhao, Vikash Kumar, Sergey Levine and [Chelsea Finn](https://aiwiki.ai/wiki/chelsea_finn), posted on 23 April 2023 [25][26]. Two small leader arms are backdriven by the operator's hands, two follower arms mirror them, and the whole rig fits a $20,000 budget using Trossen's WidowX leader arms and ViperX follower arms [26][29]. On top of that data the authors introduced Action Chunking with Transformers ([ACT](https://aiwiki.ai/wiki/action_chunking_transformer)), which predicts a sequence of future actions rather than one, and reported 80 to 90 percent success on tasks such as opening a translucent condiment cup and slotting a battery from about 10 minutes of demonstrations [25].

[Mobile ALOHA](https://aiwiki.ai/wiki/mobile_aloha), from Zipeng Fu, Tony Z. Zhao and Chelsea Finn on 4 January 2024, added a wheeled base and a whole-body teleoperation interface. With 50 demonstrations per task, co-training on existing static ALOHA data raised success rates by up to 90 percent, enough for the robot to saute and serve shrimp, open a two-door cabinet, call and ride an elevator, and rinse a pan under a tap [27]. [ALOHA 2](https://aiwiki.ai/wiki/aloha_2), from a team at Google DeepMind, Stanford and Hoku Labs, reworked the leader and follower grippers with a low-friction rail, replaced rubber-band gravity compensation with a passive mechanism, swapped the webcams for Intel RealSense D405 depth cameras, and released the hardware designs and a MuJoCo simulation model [28].

A second family of interfaces removes the robot from data collection entirely, trading fidelity for throughput.

| System | Date | Interface | What it yields |
| --- | --- | --- | --- |
| ALOHA [25] | Apr 2023 | Two leader arms puppeteering two follower arms | Bimanual joint trajectories on the target robot |
| Mobile ALOHA [27] | Jan 2024 | ALOHA arms plus an operator-pushed mobile base | Whole-body mobile manipulation demonstrations |
| UMI [30] | Feb 2024 | Hand-held gripper with wrist camera, no robot present | Hardware-agnostic policies deployable across platforms |
| DexCap [31] | Mar 2024 | Portable hand mocap using SLAM and electromagnetic tracking | Human hand and finger motion retargeted by inverse kinematics |
| Open-TeleVision [32] | Jul 2024 | VR headset with stereoscopic active vision, arm and hand mirroring | Long-horizon humanoid demonstrations |
| COBALT [40] | May 2026 | Cloud teleoperation from smartphones, VR headsets, 3D mice | Crowdsourced simulation and real-world demonstrations |

The Universal Manipulation Interface (UMI), from Cheng Chi, Zhenjia Xu, Chuer Pan, Eric Cousineau, Benjamin Burchfiel, Siyuan Feng, Russ Tedrake and Shuran Song, uses hand-held grippers with careful interface design, then closes the gap to the robot with inference-time latency matching and a relative trajectory action representation, so the learned policies are hardware-agnostic [30]. DexCap, from Chen Wang, Haochen Shi, Weizhuo Wang, Ruohan Zhang, Li Fei-Fei and C. Karen Liu, captures wrist and finger motion with SLAM plus electromagnetic tracking for occlusion resistance, retargets it onto a [dexterous hand](https://aiwiki.ai/wiki/dexterous_hand) by inverse kinematics, and adds an optional human-in-the-loop correction during policy rollouts [31]. Open-TeleVision, from Xuxin Cheng, Jialong Li, Shiqi Yang, Ge Yang and Xiaolong Wang, gives the operator active stereoscopic vision through a [virtual reality](https://aiwiki.ai/wiki/virtual_reality) headset while mirroring arm and hand motion onto two different humanoid robots [32].

Aggregating this data is a field in itself. [Open X-Embodiment](https://aiwiki.ai/wiki/open_x_embodiment), released on 13 October 2023, pooled 60 existing datasets covering 22 robot embodiments and 527 skills across more than a million real robot trajectories, and showed positive transfer when a single policy was trained across them [33][34]. [AgiBot](https://aiwiki.ai/wiki/agibot)'s AgiBot World Colosseo, published on 9 March 2025, reported over a million trajectories across 217 tasks in five deployment scenarios, collected through a standardized pipeline with human-in-the-loop verification; policies pre-trained on it beat Open X-Embodiment pre-training by an average of 30 percent [35]. [Generalist AI](https://aiwiki.ai/wiki/generalist_ai) said on 4 November 2025 that its GEN-0 model was pretrained on 270,000 hours of real-world manipulation data gathered from thousands of collection devices and robots in homes, warehouses and workplaces, and that the pile was growing at 10,000 hours a week [38].

Because human hours are the scarce input, several groups now multiply them synthetically. NVIDIA's synthetic motion generation blueprint for [Isaac GR00T](https://aiwiki.ai/wiki/isaac_gr00t), described on 18 March 2025, takes a small number of human demonstrations and expands them: 780,000 [synthetic](https://aiwiki.ai/wiki/synthetic_data) trajectories generated in 11 hours, which NVIDIA equates to 6,500 hours (nine continuous months) of human demonstration data, improving [GR00T N1](https://aiwiki.ai/wiki/groot_n1) performance by 40 percent compared with real data alone [39]. The blueprint's teleoperation front end streams Isaac Lab simulations to an [Apple Vision Pro](https://aiwiki.ai/wiki/apple_vision_pro) headset through NVIDIA CloudXR Runtime and takes control data back, but that component was listed as coming soon when the pipeline was published, with a space mouse used to record motions in the meantime [39]. Crowdsourcing is the other lever. COBALT, published in May 2026, runs vectorized environments so that dozens of remote users can teleoperate concurrently at 20 Hz with sub-100 millisecond end to end latency, up to eight users per GPU and 256 simulated clients across eight GPUs, and its pilot dataset gathered over 7,500 demonstrations, more than 50 hours, from operators in nine countries in five days [40].

## Humanoids and the autonomy question

Because teleoperation and autonomy produce visually identical demonstrations, humanoid robot videos have become a recurring source of dispute.

Tesla's We, Robot event on 10 October 2024 at Warner Bros. Studios Burbank is the standard example. [Tesla Optimus](https://aiwiki.ai/wiki/tesla_optimus) units walked without external control, but employees stationed remotely oversaw many of the interactions between the robots and attendees, and video from the event showed an Optimus working as a bartender acknowledging that it was being "assisted by a human" [41][42]. The bots all had different voices, and their responses and hand gestures were immediate and synchronized, which was the giveaway [41].

[Figure AI](https://aiwiki.ai/wiki/figure_ai) has taken the opposite tack, stating autonomy explicitly and documenting where the teleoperation actually sits. Its [Helix](https://aiwiki.ai/wiki/figure_helix) [vision-language-action model](https://aiwiki.ai/wiki/vision_language_action_model), announced 20 February 2025, was trained on roughly 500 hours of a multi-robot, multi-operator dataset of teleoperated behaviors, auto-labeled by a vision language model asked what instruction would have produced each clip; at run time an onboard VLM runs at 7 to 9 Hz and a reactive visuomotor policy emits actions at 200 Hz [36]. The successor, Helix 02, published 27 January 2026, cites over 1,000 hours of joint-level retargeted human motion data and states that the demonstration videos are "fully autonomous, not teleoperated" [37].

Consumer humanoids make the tension commercial. [1X Technologies](https://aiwiki.ai/wiki/1x_technologies) opened pre-orders for its [NEO](https://aiwiki.ai/wiki/1x_neo) home robot on 28 October 2025 at $20,000 outright or $499 a month, with United States deliveries scheduled for 2026 [44][45]. 1X states that "for complex tasks NEO doesn't know, an Expert from 1X can remotely supervise its actions at scheduled times to help it learn new abilities," and offers owners the ability to pilot the robot themselves from anywhere in the world through the NEO mobile app and a VR device [43]. The Wall Street Journal's Joanna Stern reported that she did not see NEO do anything autonomously, and that a human pilot directed every action, while the company presents operator recordings as training data in a flywheel where shared data improves autonomy over time [44]. The stated privacy controls are that operators cannot connect without explicit user approval, cannot see people (who can be blurred), and are blocked in software from user-defined no-go zones [44]. That arrangement puts a stranger's [human in the loop](https://aiwiki.ai/wiki/human_in_the_loop) inside a private home, which is a labor and privacy question as much as an engineering one.

## Limitations

Delay remains the governing constraint, and Ferrell's 1965 result still describes what operators do about it: past a certain lag they stop moving continuously and fall back on move-and-wait, which is safe and very slow [3]. The control-theoretic fixes hold the loop stable across a delayed channel by constraining how energy crosses it [5][6], and predictive displays let the operator work against a simulated model rather than a late video feed [10]. A predictive display is only as good as its model, though, and the model is least reliable exactly when the remote scene does something unexpected.

Throughput is the constraint that matters for [embodied AI](https://aiwiki.ai/wiki/embodied_ai). Classical teleoperation is one operator to one robot in real time, so demonstration data accumulates at wall-clock speed and costs human wages. That is why recent work has produced robot-free capture rigs [30][31], cloud-shared operators [40] and synthetic expansion of small demonstration sets [39], all of them attempts to break the linear relationship between human hours and data.

Data quality is a separate problem from data volume. The ACT paper's motivation was that human demonstrations are non-stationary and policy errors compound over time, so naive [behavioral cloning](https://aiwiki.ai/wiki/behavioral_cloning) on teleoperated data degrades quickly on precision tasks [25]. Interfaces that bypass the robot introduce an embodiment gap instead: a human hand recorded by motion capture does not have the robot's kinematics, so the retargeting step becomes a source of error [31].

There is also a presentation problem. A teleoperated demonstration and an autonomous one look the same on video, and viewers cannot tell them apart unless the developer says which is which [41]. Figure labels its Helix 02 videos as fully autonomous and not teleoperated [37]; Tesla attached no such label to the We, Robot demonstrations, and the human involvement surfaced through press reporting afterwards [41].

## See also

- [Robot teleoperation](https://aiwiki.ai/wiki/robot_teleoperation)
- [Imitation learning](https://aiwiki.ai/wiki/imitation_learning)
- [Behavioral cloning](https://aiwiki.ai/wiki/behavioral_cloning)
- [Mobile ALOHA](https://aiwiki.ai/wiki/mobile_aloha)
- [Surgical robot](https://aiwiki.ai/wiki/surgical_robot)
- [Humanoid robot](https://aiwiki.ai/wiki/humanoid_robot)

## References

1. "Raymond Goertz." Wikipedia. https://en.wikipedia.org/wiki/Raymond_Goertz
2. Goertz, Raymond C. "Remote-control manipulator." U.S. Patent 2,632,574, filed 16 December 1949, granted 24 March 1953. https://patents.google.com/patent/US2632574A/en
3. Ferrell, William R. "Remote manipulation with transmission delay." IEEE Transactions on Human Factors in Electronics HFE-6, no. 1 (1965): 24-32. https://doi.org/10.1109/THFE.1965.6591253
4. Sheridan, Thomas B. Telerobotics, Automation, and Human Supervisory Control. Cambridge, MA: MIT Press, 1992. ISBN 9780262193160. https://openlibrary.org/works/OL4274944W
5. Anderson, Robert J., and Mark W. Spong. "Bilateral control of teleoperators with time delay." IEEE Transactions on Automatic Control 34, no. 5 (1989): 494-501. https://doi.org/10.1109/9.24201
6. Niemeyer, G., and J.-J. E. Slotine. "Stable adaptive teleoperation." IEEE Journal of Oceanic Engineering 16, no. 1 (1991): 152-162. https://doi.org/10.1109/48.64895
7. Massie, Thomas H., and J. K. Salisbury. "The PHANTOM Haptic Interface: A Device for Probing Virtual Objects." Proceedings of the ASME Winter Annual Meeting, Symposium on Haptic Interfaces for Virtual Environment and Teleoperator Systems, Chicago, November 1994. https://ai.stanford.edu/~jks/pubs/1994-Massie-salisbury-asme-sohi.pdf
8. "Lunokhod 1." Wikipedia. https://en.wikipedia.org/wiki/Lunokhod_1
9. "Lunokhod programme." Wikipedia. https://en.wikipedia.org/wiki/Lunokhod_programme
10. "ROTEX (1988-1993)." German Aerospace Center (DLR), Institute of Robotics and Mechatronics. https://www.dlr.de/en/rm/research/robotic-systems/history/rotex-1988-1993
11. "STS-55." Wikipedia. https://en.wikipedia.org/wiki/STS-55
12. "Robonaut 2." NASA. https://www.nasa.gov/robonaut2/
13. "Robonaut." Wikipedia. https://en.wikipedia.org/wiki/Robonaut
14. "ESA-ISS experiment: Astronaut steers rover on Earth." German Aerospace Center (DLR), 2019. https://www.dlr.de/en/rm/latest/news/2019/esa-iss-experiment-astronaut-steers-rover-on-earth
15. "Jason (ROV)." Wikipedia. https://en.wikipedia.org/wiki/Jason_(ROV)
16. "Nereus (underwater vehicle)." Wikipedia. https://en.wikipedia.org/wiki/Nereus_(underwater_vehicle)
17. "Fuel debris sample extracted from damaged Japanese reactor." World Nuclear News, 7 November 2024. https://www.world-nuclear-news.org/articles/Fuel-debris-sample-extracted-from-damaged-Japanese-reactor
18. "Da Vinci Surgical System." Wikipedia. https://en.wikipedia.org/wiki/Da_Vinci_Surgical_System
19. Gottlieb, Scott. "Surgeons perform transatlantic operation using fibreoptics." BMJ, 29 September 2001. https://pmc.ncbi.nlm.nih.gov/articles/PMC1121281/
20. Marescaux, J., J. Leroy, M. Gagner, F. Rubino, D. Mutter, M. Vix, S. E. Butner and M. K. Smith. "Transatlantic robot-assisted telesurgery." Nature 413(6854): 379-380, 27 September 2001. https://pubmed.ncbi.nlm.nih.gov/11574874/
21. "Intuitive da Vinci 5 updates offer feedback during surgery." MedTech Dive, 15 September 2025. https://www.medtechdive.com/news/Intuitive-Surgical-da-Vinci-new-software-features-force-feedback/760069/
22. "da Vinci 5." Intuitive Surgical. https://www.intuitive.com/en-us/products-and-services/da-vinci/5
23. "Fleet Response: lending a helpful hand to Waymo's autonomously driven vehicles." Waymo, 21 May 2024. https://waymo.com/blog/2024/05/fleet-response/
24. "Vay launches commercial driverless mobility service with remotely driven cars in Las Vegas, Nevada." Vay, 17 January 2024. https://vay.io/vay-launches-commercial-driverless-mobility-service-with-remotely-driven-cars-in-las-vegas-nevada/
25. Zhao, Tony Z., Vikash Kumar, Sergey Levine and Chelsea Finn. "Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware." arXiv:2304.13705, 23 April 2023. https://arxiv.org/abs/2304.13705
26. "ALOHA: A Low-cost Open-source Hardware System for Bimanual Teleoperation." Project page. https://tonyzhaozh.github.io/aloha/
27. Fu, Zipeng, Tony Z. Zhao and Chelsea Finn. "Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation." arXiv:2401.02117, 4 January 2024. https://arxiv.org/abs/2401.02117
28. "ALOHA 2: An Enhanced Low-Cost Hardware for Bimanual Teleoperation." Project page. https://aloha-2.github.io/
29. "Aloha Stationary." Trossen Robotics. https://www.trossenrobotics.com/aloha-stationary
30. Chi, Cheng, Zhenjia Xu, Chuer Pan, Eric Cousineau, Benjamin Burchfiel, Siyuan Feng, Russ Tedrake and Shuran Song. "Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots." arXiv:2402.10329, 15 February 2024. https://arxiv.org/abs/2402.10329
31. Wang, Chen, Haochen Shi, Weizhuo Wang, Ruohan Zhang, Li Fei-Fei and C. Karen Liu. "DexCap: Scalable and Portable Mocap Data Collection System for Dexterous Manipulation." arXiv:2403.07788, 12 March 2024. https://arxiv.org/abs/2403.07788
32. Cheng, Xuxin, Jialong Li, Shiqi Yang, Ge Yang and Xiaolong Wang. "Open-TeleVision: Teleoperation with Immersive Active Visual Feedback." arXiv:2407.01512, 1 July 2024. https://arxiv.org/abs/2407.01512
33. Open X-Embodiment Collaboration. "Open X-Embodiment: Robotic Learning Datasets and RT-X Models." arXiv:2310.08864, 13 October 2023. https://arxiv.org/abs/2310.08864
34. "Open X-Embodiment: Robotic Learning Datasets and RT-X Models." Project page. https://robotics-transformer-x.github.io/
35. AgiBot-World-Contributors et al. "AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Systems." arXiv:2503.06669, 9 March 2025. https://arxiv.org/abs/2503.06669
36. "Helix: A Vision-Language-Action Model for Generalist Humanoid Control." Figure, 20 February 2025. https://www.figure.ai/news/helix
37. "Introducing Helix 02: Full-Body Autonomy." Figure, 27 January 2026. https://www.figure.ai/news/helix-02
38. "GEN-0." Generalist AI, 4 November 2025. https://generalistai.com/blog/gen-0
39. "Building a Synthetic Motion Generation Pipeline for Humanoid Robot Learning." NVIDIA Developer Blog, 18 March 2025. https://developer.nvidia.com/blog/building-a-synthetic-motion-generation-pipeline-for-humanoid-robot-learning/
40. Agarwal, Ayush, Ansh Gandhi, Jeremy A. Collins, Omar Rayyan, Aryan Sarswat, Ranjani Koushik, Masoud Moghani, Ajay Mandlekar and Animesh Garg. "COBALT: Crowdsourcing Robot Learning via Cloud-Based Teleoperation with Smartphones." arXiv:2605.19138, 18 May 2026. https://arxiv.org/abs/2605.19138
41. Wiggers, Kyle. "Tesla's Optimus bots were controlled by humans during the We, Robot event." TechCrunch, 14 October 2024. https://techcrunch.com/2024/10/14/tesla-optimus-bots-were-controlled-by-humans-during-the-we-robot-event/
42. "Tesla Cybercab." Wikipedia. https://en.wikipedia.org/wiki/Tesla_Cybercab
43. "NEO Home Robot." 1X Technologies. https://www.1x.tech/neo
44. "1X Details NEO's 'Human-in-the-Loop' Strategy and Hardware as Pre-Orders Go Live." Humanoids Daily, 28 October 2025. https://www.humanoidsdaily.com/news/1x-details-neo-human-in-the-loop-strategy-and-hardware-as-pre-orders-go-live
45. "Order NEO." 1X Technologies. https://www.1x.tech/order
46. Goertz, R. C., and W. M. Thompson. "Electronically Controlled Manipulator." Nucleonics 12, no. 11 (November 1954). https://www.osti.gov/biblio/4387399
47. "Remote surgery." Wikipedia. https://en.wikipedia.org/wiki/Remote_surgery