On this page
Data and Teleoperation
In 2024, Tesla demonstrated its Optimus robot folding a shirt on stage at the "We, Robot" event. The movements looked smooth, almost effortless. What...
In 2024, Tesla demonstrated its Optimus robot folding a shirt on stage at the "We, Robot" event. The movements looked smooth, almost effortless. What most viewers didn't know was that this seemingly simple task represented thousands of hours of training, not in the physical world, but in simulation. And simulation, it turns out, has limits.
For decades, robotics companies have relied on simulated environments to train their machines. It's efficient: you can run thousands of virtual trials in the time it takes to run one real test. You can create perfect conditions, eliminate variables, and iterate rapidly without worrying about broken hardware or safety hazards. But there's a problem. The real world refuses to behave like a simulation.
A shirt in simulation has predictable fabric physics. A real shirt wrinkles unpredictably, slips through fingers, and behaves differently when damp versus dry. A simulated factory floor is clean and organized. A real factory has oil spills, workers walking unexpected paths, and lighting that changes throughout the day. This gap between simulation and reality, what researchers call the "sim-to-real transfer problem," is why humanoid robots need something simulation cannot provide: messy, unpredictable, real-world data.
The Data Feedback Loop#
Figure AI deployed its Figure 02 robot at BMW's Spartanburg plant in South Carolina starting in August 2024. As of March 2025, a single Figure 02 robot was operating during production hours, performing sheet metal insertion tasks in the body shop. The task itself is straightforward: pick up metal parts from a container and place them into fixtures with precision tighter than one centimeter.
But here's what makes this deployment valuable: every placement generates data. By November 2024, Figure announced the robot was performing up to 1,000 placements per day, achieving a 400% increase in speed and a sevenfold improvement in success rate compared to earlier tests. Each attempt teaches the AI something about grip force, part orientation, fixture alignment, lighting conditions, and the million small variables that distinguish a real factory from a simulated one.
"This will only improve as we deploy more robots, collect more data, and improve our AI models," stated Brett Adcock, Figure's CEO. This is the data feedback loop in action: deploy robots, collect performance data, improve AI models, deploy better robots. The more robots in the field, the faster the learning accelerates.
China's Factory Floor Laboratory#
While American companies test robots in pilot programs, Chinese manufacturers are treating entire factory floors as data collection laboratories. UBTECH's Walker S1 humanoid robots are deployed in multiple automotive facilities including Audi-FAW, Zeekr, BYD, Geely, and Foxconn, with over 500 units pre-ordered.
At Zeekr's 5G-enabled smart factory in Ningbo, dozens of Walker S1 robots work collaboratively in quality inspection, vehicle assembly, parts sorting, and final assembly. These robots are connected through UBTECH's "BrainNet" framework, a distributed AI system that enables swarm intelligence. When one robot learns a more efficient way to grasp a deformable material or navigate around an obstacle, that knowledge propagates across the fleet.
At BYD's Shenzhen plant, Walker S1 robots increased sorting efficiency by 120 percent. Each efficiency gain represents data: sensor readings about object weight and texture, motor feedback about optimal grip force, vision data about lighting and angles. This data trains the next generation of AI models, which makes the next wave of robots more capable.
The strategy is clear: get robots into diverse real-world environments as quickly as possible, collect massive amounts of operational data, and iterate rapidly. UBTECH announced plans to begin mass production of Walker S humanoid robots by year-end 2025, targeting delivery of 500 to 1,000 units.
Why Data Cannot Be Shared#
Real-world humanoid data has a peculiar characteristic: it doesn't transfer well. Data collected in a BMW factory teaches robots to handle BMW's specific parts, fixtures, lighting, and workflow. It won't directly help a robot working in an Amazon warehouse or a hospital. Data collected in California factories differs from data collected in Chinese factories because temperature, humidity, worker behaviors, and material handling practices differ.
This context-specificity means that even as the humanoid robotics industry matures, data remains fragmented. Companies like Audi, BYD, and Zeekr collect proprietary data through their Walker S1 deployments. Amazon collects data through its warehouse robots. Each dataset represents competitive advantage in a specific domain.
Regulatory barriers reinforce this fragmentation. European companies operating under GDPR face restrictions on data collection and sharing, particularly when robots operate in spaces where they might capture images of workers. China's data localization laws require that data collected within China remain on Chinese servers. These regulations don't just slow data sharing; they make it legally complex or impossible.
The result is a robotics landscape where no single company can build a universal humanoid AI model trained on all possible real-world scenarios. Instead, we're heading toward specialization: robots optimized for automotive manufacturing, robots optimized for warehousing, robots optimized for eldercare, each trained on data specific to their domain.
Teleoperation: The Bridge Technology#
The first humanoid robots to enter our homes will not be fully autonomous. Neither will many of those deployed in factories. This is not a failure of technology. It is the reality of how complex machines learn to operate in an unpredictable world.
In October 2025, 1X Technologies opened pre-orders for NEO, a humanoid robot designed for homes, at $20,000 upfront or $499 per month. The robot could perform basic autonomous tasks like opening doors, fetching items, and turning lights on or off. But for anything more complex, anything specific to a particular home or family's needs, a human teleoperator would take control, viewing the inside of buyers' homes through the robot's cameras and manipulating its limbs remotely.
CEO Bernt Børnich was remarkably candid about the arrangement: "If we don't have your data, we can't make the product better." Owners would schedule sessions through an app, specifying tasks they wanted completed, and remote operators would execute them while the robot's AI systems watched and learned. Early adopters were not buying autonomous assistants. They were funding the training ground for future autonomy.
This is teleoperation, the bridge technology that makes humanoids useful today while building toward autonomy tomorrow. A human operator wears specialized equipment, a VR headset and haptic gloves or motion capture suit, that transmits their movements to the robot in real time. The robot's sensors and cameras send information back, creating an immersive experience of embodiment from afar. The operator sees through the robot's eyes, feels resistance when grasping objects, and controls movements as naturally as their own body.
Sanctuary AI has built its entire development strategy around this approach, using what it calls "analogous teleoperation" where pilots wearing specialized rigs control Phoenix humanoid robots to complete complex manipulation tasks. During a week-long 2023 pilot at a Mark's retail store in British Columbia, a teleoperated Phoenix successfully completed 110 different retail-related tasks. Every action was recorded, creating datasets for autonomous systems.
The results are measurable. Sanctuary AI reported that the time required to automate new tasks dropped from weeks to less than 24 hours between their sixth and seventh generation Phoenix robots. 1X demonstrated robots chaining together multiple skills via voice commands, with operators training both low-level motor skills and high-level task sequences. Each teleoperation session accelerates the path toward machines that can work independently.
A Third Path: Capturing Human Skills#
Not everyone agrees that teleoperation is the answer. In November 2025, a startup called Sunday Robotics emerged from stealth with $35 million in funding and a radically different approach to the data problem.
The company was founded by Stanford PhD roboticists Tony Zhao and Cheng Chi. Zhao had previously interned at DeepMind, Tesla Autopilot, and Google X. When Sunday unveiled its home robot Memo, the technical specifications were interesting but not exceptional: a wheeled semi-humanoid standing 1.7 meters tall with 20 degrees of freedom, designed for household chores like dishes, laundry, and tidying. What set Sunday apart was how Memo learned.
Instead of teleoperation, Sunday developed a $200 wearable called the Skill Capture Glove. The company distributed these gloves to hundreds of ordinary people, whom they call "Memory Developers." These participants wear the gloves while doing chores in their own homes, recording how they move, clean, and organize. The glove captures hand movements, grip patterns, and task sequences. Sunday then transfers this data directly to Memo's AI, training the robot on human demonstrations without ever needing a human to control a robot.
"The problem has always been data," Zhao explained at the launch. "Most home robots start as adaptations of industrial machines, and those trained in labs rarely succeed in unpredictable, real-world environments. Our Skill Capture Glove changes this by collecting thousands of hours of daily routines from hundreds of families."
The scale advantage is significant. By late 2025, Sunday had collected approximately 10 million episodes of genuine household routines from over 500 real homes. Compare this to teleoperation, where every training session requires an expensive robot, a trained operator, and real-time coordination. Sunday's approach decouples data collection from robot deployment entirely. You don't need robots to gather robot training data. You just need gloves and willing participants.
The contrast with 1X Technologies is instructive. When 1X launched NEO, buyers became data sources: teleoperators would control their robots remotely, and those sessions would train the AI. Sunday inverted this model. Their data collection happens before any robot enters a home. By the time Memo arrives, it has already learned from millions of demonstrations across hundreds of different kitchens, living rooms, and laundry spaces.
Eric Vishria, General Partner at Benchmark, framed the stakes bluntly: "We have about one-millionth of the data we need. Tony and Cheng's approach finally makes collecting robot-ready data at a massive scale possible."
The industry is now watching three distinct philosophies compete. 1X Technologies bets on teleoperation: humans control robots, robots learn from being controlled. Figure AI takes an autonomy-first stance, refusing to teleoperate and instead solving intelligence through deployed robots gathering their own experience. Sunday Robotics offers a third path: capture human skills directly, at massive scale, without robots in the loop at all.
Which approach wins may depend on the domain. Factory robots operating in controlled environments may thrive on autonomous data collection. Home robots facing infinite variation may need the breadth that skill capture provides. And for tasks requiring real-time human judgment, teleoperation may remain essential regardless of how much training data exists.
The deeper lesson is that the data problem has no single solution. The companies that succeed will likely be those that match their data strategy to their deployment context, rather than betting everything on one approach.
The Hybrid Model#
The economics favor a hybrid approach. Pure teleoperation is not scalable. Every robot requires a human operator's time, creating a one-to-one ratio that defeats automation's purpose. Labor costs move from physical locations to remote control centers but do not disappear. However, hybrid systems that operate autonomously for learned behaviors and hand control to humans for exceptions create a viable path forward.
The robot handles routine tasks independently. Humans intervene only when necessary. As the robot's capabilities expand, the ratio shifts. More tasks become autonomous. Human supervision becomes occasional rather than constant.
This architecture will persist even as autonomy improves. There will always be edge cases, situations robots have not encountered, tasks requiring judgment that machines cannot yet replicate. The seamless handoff between autonomous operation and human control creates a continuous learning cycle. Humanoids become more capable with every intervention, approaching but perhaps never fully achieving complete independence from human guidance.
Teleoperation also functions as a safety mechanism. Even as humanoids become more autonomous, the ability for humans to take control in unexpected situations remains critical, especially in home environments where mistakes could harm people or property. When a robot encounters something outside its training, human judgment prevents failures and simultaneously teaches the system how to handle novel situations.
Beyond Earth#
The principles that make teleoperation valuable on Earth become essential beyond it. NASA and the European Space Agency have been testing teleoperation from the International Space Station, with astronauts controlling robots on Earth to simulate future Mars operations. In August 2025, NASA astronaut Jonny Kim controlled multiple robots simultaneously from the ISS, including a humanoid robot called Rollin' Justin, navigating simulated Mars terrain and performing sample collection tasks.
The concept is straightforward: before humans land on Mars or the Moon, astronauts in orbit would send robots to the surface to build habitats and infrastructure. Teleoperation from orbit solves problems that Earth-based control cannot. Radio signals take between four and twenty-four minutes to travel between Earth and Mars, depending on the planets' positions. This delay makes real-time control impossible. But an operator in Mars orbit experiences virtually no delay. Milliseconds separate command from execution.
SpaceX plans to send Tesla's Optimus humanoid robot to Mars by the end of 2026, with Elon Musk stating the robots would prepare the ground for future human colonies. Whether this timeline proves realistic matters less than what it signals: humanoid robots are moving beyond factory floors into the most extreme environments imaginable. NASA's Valkyrie robot, developed at Johnson Space Center starting in 2015, was explicitly designed for future Mars missions, built to precede humans and prepare habitats.
And if we can teleoperate robots on Mars, we can teleoperate them anywhere. On asteroids where mining operations require human judgment but human presence would be suicidal. In the deep ocean where pressure prevents direct human exploration. In nuclear accident sites, chemical spills, and collapsed mines. Anywhere that combines the need for human intelligence with conditions hostile to human biology.
The Implications#
Real-world data is becoming the decisive factor separating robotics demonstrations from practical deployment. Simulations can prototype ideas, but only physical robots operating in real environments generate the corner cases, failure modes, and subtle variables that train robust AI systems.
The companies moving fastest from prototypes to deployment, even in limited numbers, are building data advantages that competitors cannot easily replicate. In humanoid robotics, data collection is not a byproduct of deployment. It is the primary goal. Every robot in the field is a sensor collecting training data. Every task completed feeds the AI models that will power the next generation.
And because this data is context-specific, geographically constrained, and often proprietary, the robotics industry is evolving toward specialized niches rather than universal platforms. The age of the general-purpose humanoid robot may be arriving, but the data powering these machines will remain stubbornly specific, local, and fragmented. That fragmentation, paradoxically, may be what enables the industry's growth: no monopoly, just specialization and competition across countless domains.
References: