Table of Contents
Every robotics company is chasing the same destination: a robot that operates fully on its own, handles any situation, and never needs a human in the loop. Recently, Figure AI shared a video of their Felix 02 'fully autonomous' humanoid doing human tasks. Whilst it appears to have a huge degree of autonomy, edge cases such as difficulties with lighting mean that they will still occasionally need teleop. Figure AI are leading the race in getting to full autonomy, and would have required countless hours of teleoperation to get them to this position.
For those far behind Figure AI, robots need to be deployed. Customers need to be served. Revenue needs to be generated. Data needs to be collected.
That's why the most successful robotics companies aren't waiting for full autonomy. They're using teleop, not as a fallback, but as a deliberate strategy.
The Autonomy Gap Is Real — And Bigger Than Most Demos Suggest
If you watched the Figure AI video, you might be forgiven for thinking full autonomy is just around the corner. It isn't.
Bain & Company's 2025 Technology Report put it plainly: most humanoid robots today remain in pilot phases, heavily dependent on human input for navigation, dexterity, or task switching. "This 'autonomy gap' is real. Current demos often mask technical constraints through staged environments or remote supervision."
Even Tesla's Optimus program, one of the most heavily publicised in the industry, tells the same story. As of late 2025, Tesla had deployed "at least two" Optimus units performing tasks in its Fremont factory. Evidence from multiple sources suggests continued reliance on teleoperation during public demonstrations, with human operators remotely controlling the robots to execute impressive-looking tasks.
There are two of key stubborn technical barriers:
Battery life. Most current humanoid robots run for 90 minutes to two hours before needing a charge. An industrial deployment requires a full eight-hour shift. Agility Robotics' Digit, arguably the most production-hardened humanoid on the market after its Amazon warehouse pilots, runs for 90 minutes followed by a 9-minute fast charge. That's nowhere near shift-length operation without intervention.
Reliability standards. Traditional industrial robots achieve 95–99% uptime in well-maintained environments. Humanoid robots are nowhere near that benchmark. A humanoid has hundreds of joints, actuators, and sensors, each one a potential failure point. An industrial arm has six joints. Meanwhile, no ISO standard yet exists for dynamically balancing legged robots in human workspaces. Until those standards are finalised, enterprise procurement at scale will be limited.
What the AV Industry Already Learned
Robotics companies navigating the teleoperation-vs-autonomy question don't need to start from scratch. Autonomous vehicles went through this exact debate, and the lessons are directly applicable.
The prevailing assumption a decade ago was that human operators were a temporary crutch, to be phased out quickly as AI matured. What has actually happened was more nuanced.
Waymo, the most operationally advanced AV company in the world, keeps a team of remote human agents on standby, not to drive the cars, but to provide high-level guidance when vehicles encounter situations they aren't confident handling: unusual construction zones, police directing traffic with hand signals, ambiguous road conditions. The Waymo Driver handles the vast majority of situations independently. But the human safety net is always there, and the data from every intervention feeds back into the training pipeline.
The numbers are striking. According to California's 2024 disengagement data, the most closely watched metric in the sector, Waymo recorded one disengagement every 9,793 miles. Newer operators needed interventions far more frequently. The lesson isn't that human operators are a sign of weakness. It's that the ratio of robots to operators improves continuously as AI matures, and the operators themselves are what make deployment safe enough to generate the data that drives that improvement.
Cruise's experience offers the cautionary version of the same lesson. At the height of its operations, Cruise's vehicles needed remote human intervention approximately every four to five miles, a fact obscured in public communications. When that gap between claimed capability and operational reality became visible, it contributed to the company's collapse.
The parallel for humanoid robotics is exact. Companies that are transparent about their teleoperation dependency and treat it as an asset, as a way to generate training data while delivering real value, are building a durable path to autonomy. Companies that obscure it are building a credibility problem.
Teleoperation Isn't a Step Backward. It's the Pipeline.
Teleoperation is not a workaround for the absence of autonomy. It is how you build autonomy.
Every time a trained human operator guides a robot through a task, that interaction generates high-fidelity training data: camera video, sensor readings, force data, timestamped controls. The robot is learning what a human expert does in a real, messy, uncontrolled environment, not a simulation. This is the kind of data that foundation models need to generalise.
Sanctuary AI takes a similar approach, using remote experts alongside tactile sensors to give operators a sense of touch, generating richer training data for manipulation tasks while simultaneously solving real customer problems.
This is the architecture that works: teleoperation as a business model today, teleoperation as a data pipeline for tomorrow.
The Spectrum: Where Your Robot Actually Lives
It's tempting to think of teleoperation and autonomy as two states, on or off. In practice, every deployed robot sits somewhere on a spectrum, and the right position on that spectrum depends on your use case, your environment, and where you are in your autonomy development.
Full teleoperation: a human controls every movement. This makes sense for genuinely novel tasks, dangerous environments, or early deployments where edge cases are frequent and unpredictable.
Supervised autonomy: The robot handles routine execution while a human monitors and can intervene, is where most of the most successful deployments operate today. The robot does the work; the human provides a safety net and steps in for edge cases. Waymo's fleet response model is the gold standard here: approximately 70 remote assistants overseeing a fleet of around 3,000 vehicles.
Task-limited autonomy: full independence for specific, well-defined, repeatable tasks in controlled environments, is commercially real today. Agility's Digit doing pick-and-place in an Amazon warehouse. UBTech's Walker S2 performing repetitive manufacturing tasks on structured production lines. The key word is controlled. These robots aren't general-purpose; they're reliable in their lane.
Full autonomy: true independence across diverse, unstructured, unpredictable environments, is the destination. It is not 2026's reality for most applications. Goldman Sachs estimates that consumer applications lag industrial ones by two to four years.
The strategic question for any robotics company isn't "when will we reach full autonomy?" It's: "What's the right level of human-in-the-loop for our deployment right now, and how do we use that to accelerate toward greater autonomy over time?"
The Business Case for Getting This Right
Deploying robots that attempt full autonomy before they're ready produces visible, public failures. Those failures erode customer trust, generate negative press, and, as Cruise discovered, can be existential. The AV sector spent years overpromising on autonomy timelines. The robotics sector is at risk of repeating the same mistake.
Deploying robots with the right level of teleoperation support does something different. It lets you:
- Generate revenue today, even before full autonomy is achieved
- Collect real-world training data at scale, in the environments your robots will eventually operate autonomously, allowing for labelling with platforms such as labellerr.
- Build customer trust through reliable performance, rather than eroding it through visible failures
- Reduce deployment risk by keeping humans in the loop for the edge cases that will inevitably arise
- Create a defensible moat — the operator expertise, training data, and operational playbooks you build now are hard for competitors to replicate
The teleoperation market itself reflects this shift. It was valued at approximately $502 million in 2024 and is projected to reach $4.7 billion by 2035, growing at a compound annual rate of over 25%. That growth isn't because autonomy is failing. It's because the industry has accepted that the path to autonomy runs through teleoperation, not around it.
The Bottom Line
Full autonomy is the right destination for the robotics industry. But the path there isn't a straight line, and the companies that will get there fastest aren't the ones waiting at the starting line until their AI is perfect.
They're the ones deploying now, learning from real-world data, and using teleoperation as an engine, generating value for customers today while building the training foundation for tomorrow.
The question for robotics companies shouldn't be whether to use teleoperation. It should be whether you're using it strategically. Adamo ensures the answer is yes.
