Teleoperation and the Data Bottleneck: How Humanoid Companies Actually Collect Training Data

Ask any humanoid robotics engineer what’s holding back general-purpose robots, and the answer is rarely “the arms” or “the chips.” It’s data — specifically, the labeled, physical, embodied demonstrations that teach a robot’s neural network how to fold a towel, pick up a shrimp, or navigate a cluttered hallway. Unlike large language models, which can hoover up trillions of tokens already sitting on the internet, robots need examples of actions in the physical world, and almost none of that exists yet. So companies build it themselves, by hand, using teleoperation: a human physically drives a robot’s body through a task, frame by frame, while it records everything.

What is teleoperation, and why is it the default data source for humanoid AI?

Teleoperation means a person remotely or directly controls a robot’s movements — often through a matching exoskeleton, VR headset, joystick rig, or handheld controllers — while the robot’s sensors record the resulting joint angles, camera frames, and force feedback as a labeled demonstration. That recorded trajectory becomes training data for imitation learning, the technique where a neural network learns to reproduce a behavior by watching (and eventually mimicking) many examples of a human doing it correctly.

The reason teleoperation dominates isn’t that it’s elegant. It’s that nothing else currently produces embodied training data at usable quality. Simulation can generate infinite synthetic episodes, but transferring a policy trained purely in simulation to a real robot’s actual motors, friction, and sensor noise remains unreliable — a persistent challenge covered in more depth in our explainer on sim-to-real transfer and digital twins. Pure video scraping (watching YouTube cooking videos, for instance) gives a model visual context but no ground-truth record of the exact joint torques and gripper forces needed to actually perform the task. Teleoperation sits in between: it’s slow and expensive, but it produces exactly the paired (observation, action) data imitation learning needs.

How expensive is a teleoperation rig, really?

More expensive than most people assume, and cheaper than you’d think for what it does. Stanford’s Mobile ALOHA project — a bimanual, mobile whole-body teleoperation platform built by researchers Zipeng Fu, Tony Z. Zhao and Chelsea Finn — priced a full mobile manipulation rig, including onboard compute and power, at roughly $32,000. That system could learn tasks like cooking shrimp or opening a cabinet from around 50 human-teleoperated demonstrations, which was itself notable: a fraction of the thousands of examples earlier imitation-learning setups typically required.

The earlier, stationary ALOHA rig (the non-mobile predecessor) is documented in ALOHA pricing materials as costing roughly $27,067 in total hardware, with the robot arms and cameras alone accounting for about $20,486 — around three-quarters of the cost. It’s worth flagging that this specific breakdown comes from secondary documentation rather than an independently verified Stanford or Trossen Robotics source, so treat the exact split as directionally accurate rather than gospel. Either way, the pattern holds: teleoperation hardware sits in the tens of thousands of dollars per rig, and every rig still needs a trained human operator sitting behind it, hour after hour, to generate data.

How does Figure AI collect its training data?

Figure trained its Helix vision-language-action model on approximately 500 hours of teleoperated demonstration data, collected across multiple robots and multiple human operators. Figure has described this as less than 5% the size of previously collected VLA datasets in the field — a claim that, if accurate, suggests better data curation and an auto-labeling pipeline matter more than raw volume. An auto-labeling vision-language model generates hindsight natural-language instructions for each recorded segment, effectively turning raw teleoperation footage into labeled (video, language, action) triplets without a human writing captions by hand.

Figure has also been unusually well-capitalized to fund this kind of data operation. In February 2024 the company raised $675 million at a $2.6 billion valuation, alongside a collaboration agreement with OpenAI, with investors including Microsoft, the OpenAI Startup Fund, NVIDIA, Jeff Bezos’s Bezos Expeditions, and others. By September 16, 2025, Figure closed a Series C exceeding $1 billion at a $39 billion post-money valuation, led by Parkway Venture Capital with Brookfield Asset Management, NVIDIA, and Intel Capital participating — roughly a 15x jump in valuation in eighteen months.

That Brookfield relationship isn’t just financial. Announced alongside the Series C, “Project Go-Big” is a partnership to collect egocentric human video data across Brookfield’s real-estate holdings — reported at more than 100,000 residential units, 500 million square feet of commercial office space, and 160 million square feet of logistics space. Figure describes this as building the world’s largest humanoid pretraining dataset, and reports that Helix, trained on 100% egocentric human video with zero robot demonstrations in this phase, was able to learn to navigate cluttered home environments purely from natural-language commands. If that holds up at scale, it would mark a real shift away from teleoperation as the sole data source — from robot-generated demonstrations toward the same kind of internet-scale human video that trained today’s large language and vision models. This kind of foundation-model approach is closely tied to the broader wave of embodied foundation models discussed in our piece on NVIDIA’s Isaac and GR00T platforms.

How does 1X Technologies turn teleoperation into a product feature?

1X takes a different approach: it ships teleoperation directly into the customer experience rather than hiding it in a lab. The company’s NEO home humanoid — priced at $20,000 to buy outright or $499 per month on a six-month minimum subscription, with US and Canada consumer deliveries targeted for Q3–Q4 2026 — runs on a system 1X calls “Expert Mode.” When NEO’s onboard AI can’t complete a household task autonomously, a trained remote human operator takes over, finishes the task themselves through the robot, and that session becomes labeled training data fed back into 1X’s Redwood AI model.

This is a genuine data flywheel: every failure the robot’s autonomy stack hits becomes a data point that improves the next model iteration, and 1X has been candid about the current split. The company has stated it expects roughly 60–70% autonomous operation at initial launch, rising toward 80–90% by 2027 and 95%+ by 2028 as Redwood absorbs more of this teleoperated correction data. In other words, 1X isn’t claiming NEO is autonomous today — it’s explicitly selling a robot that leans on humans in the loop while betting the ratio improves each year.

Backing that bet has meant serious capital. In September 2025, 1X — backed by OpenAI and EQT — was reported to be in talks to raise up to $1 billion at a targeted valuation of at least $10 billion, more than 12 times its valuation from January 2024. It’s worth being precise here: multiple outlets described this as a round being sought or negotiated at the time, not one confirmed as closed, so the $10 billion figure should be read as a target rather than an announced fact.

How is Tesla’s approach to Optimus data different?

Tesla shifted its own data-collection method around mid-2025, moving away from the motion-capture suits, VR headsets, and teleoperation rigs its dedicated data-collection workforce had previously used. The newer approach has operators wear helmet- and backpack-mounted camera rigs — with some haptic gloves for hand tracking — to record everyday human tasks directly, rather than driving a robot body through them. This transition happened around the same time Milan Kovac stepped down as Optimus program lead in June 2025, with Ashok Elluswamy taking over the effort.

The distinction matters: Tesla’s newer method captures human movement in the world rather than puppeting a robot’s joints through a human operator, which is closer in spirit to Figure’s egocentric-video approach than to classic ALOHA-style teleoperation. It reflects a broader industry question that doesn’t yet have a settled answer: is it more efficient to teach a robot by literally piloting its body, or by recording humans doing tasks naturally and later solving the harder problem of retargeting that motion onto a different physical form?

Comparing the three approaches

Company	Primary data method	Notable figures	Status as of mid-2026
Figure AI	Teleoperation (Helix training) + egocentric human video (Project Go-Big)	~500 hrs teleop data for Helix; Brookfield portfolio (100K+ units, 500M sq ft office, 160M sq ft logistics) for video collection	$1B+ Series C at $39B valuation, Sept 2025
1X Technologies	Teleoperation as live product feature (“Expert Mode”) feeding the Redwood model	60–70% autonomy at launch, targeting 80–90% by 2027, 95%+ by 2028	Reportedly seeking ~$1B at $10B+ valuation (unconfirmed as closed), Sept 2025
Tesla Optimus	Shifted from motion-capture/teleoperation to camera-only human data capture	Transition dated to ~June 2025, alongside leadership change	Camera-based collection ongoing

Why is teleoperated data still the bottleneck, not compute or hardware?

Because every one of these methods still funnels through a human being — physically or virtually — performing one task at a time. A GPU cluster scales by adding racks; a data-collection operation scales by adding hours of skilled human labor, one demonstration at a time, on hardware that costs tens of thousands of dollars per rig. Two throughput figures circulate in industry commentary — that a single operator might record on the order of 20–30 episodes per hour, or alternatively that current teleoperation pipelines produce fewer than 200 demonstrations per day — but these numbers trace back to overlapping blog sources rather than independently verified company or academic data, so we’re flagging them as unconfirmed rather than treating them as fact. What is well documented is the direction every major player is moving: toward reducing how much teleoperated data a model needs per task (Mobile ALOHA’s ~50-demonstration benchmark), toward auto-labeling pipelines that cut human annotation work, and toward supplementing or replacing teleoperation with passively captured human video at much greater scale.

Where does this fit into the broader physical AI picture?

Teleoperation is essentially the manual-labor phase of a much larger production pipeline that also includes simulation, foundation models, and fleet-scale deployment — the layers mapped out in our overview of the physical AI stack. It’s the part of the stack humanoid companies are most eager to shrink, because it’s the part that scales the worst. Whether that means egocentric video at real-estate scale, simulation-first training pipelines, or simply better sample efficiency per demonstration, the throughline across Figure, 1X, and Tesla is the same: teleoperation got the industry’s humanoid robots to their current capability level, but nobody is betting their long-term roadmap on scaling it up linearly forever.

For a fuller map of how these pieces — data collection, simulation, foundation models, and fleet operations — connect, see our Physical AI Stack hub. For definitions of terms like imitation learning and vision-language-action models, check the glossary.