A humanoid robot walks across a kitchen, opens a cabinet, retrieves a mug, and sets it on the counter. The clip is shared widely. The caption reads autonomous. The reality is more complicated, and very often, a person wearing a VR headset is doing the hard part.
Teleoperation is the quiet center of gravity in humanoid robotics in 2026. Almost every public demo from Tesla Optimus, Figure 02, 1X NEO, Apptronik Apollo, and Agility Digit involves a human in the loop at some level. That is not necessarily dishonest, and in some cases the manufacturer has been transparent about it. The problem is that viewers, and many buyers, treat the demo as evidence of autonomous capability when the underlying system is not autonomous at all.
When a humanoid robot picks up a sock or folds a shirt in a viral clip, a person is often driving it. Robovations classifies teleoperated systems at Level 1 and supervised demos at Level 2, regardless of how natural the motion looks. Reading humanoid marketing critically starts with one question: who is in the loop?
What teleoperation actually means
Teleoperation in modern humanoid platforms takes several forms, and the distinctions matter for classification.
- Direct VR control: A pilot wears a headset and tracking gloves. The robot mirrors the pilot in real time. Every motion is generated by the human. The robot is essentially a remote-controlled puppet.
- Supervised autonomy: The robot performs scripted or learned behavior, but a human watches and can take over if something goes wrong. The robot may complete the task on its own, but the human is the safety net.
- Training-data capture: The pilot teleoperates the robot through a task to generate demonstrations. Those demonstrations are later used to train an autonomous policy. The robot in the demo is teleoperated; a future version may be autonomous.
- Expert-pilot fallback: The robot runs autonomously most of the time, but a remote operator can intervene for novel situations. This is closer to a video-call escalation than to autonomy.
Each of these has different implications for the Autonomy Ladder. Robovations does not treat them as interchangeable.
Why the Autonomy Ladder caps teleoperated robots at Level 1
The Autonomy Ladder measures what the robot does on its own. Level 1, Manual Automation, covers any system where a human is generating the action sequence. Direct VR teleoperation falls squarely in this tier no matter how impressive the resulting motion looks. The robot is a sophisticated mechanical extension of the human, not an autonomous agent.
Level 2, Assisted Autonomy, applies to demos where the robot executes a scripted or learned behavior in a controlled environment without human action generation, but with human supervision. The robot is doing the doing; the human is watching. Most published end-to-end humanoid demos in 2024 to 2026 land here, including Figure’s BMW manufacturing footage and 1X NEO’s household scenarios.
Level 3, Conditional Autonomy, requires the robot to complete a task end-to-end without teleoperation, in conditions that match its training distribution. This is harder than it sounds. The robot needs to handle perception, planning, and execution while recovering from minor variations on its own.
Reading humanoid demos critically
The most useful question to ask of any humanoid demo is what was end-to-end and what was edited or assisted. Concrete questions worth asking:
- Was a human teleoperating the robot during the demo?
- If autonomous, how many takes were used to produce the clip?
- What is the success rate of the same behavior across attempts?
- Was the environment a working setting or a controlled stage?
- Did the robot recover from any errors during the take, or was the clip restarted?
Demonstrations rarely include this information voluntarily. When a manufacturer publishes a clip without a success rate, an environment description, or a take count, the conservative interpretation is that the conditions favored the robot.
How disclosure varies across manufacturers
1X has documented publicly about the role of teleoperation. The company has stated that NEO Beta and Gamma include both autonomous and teleoperated modes, and that human pilots can take over for tasks the robot cannot do alone. The household subscription model explicitly contemplates expert-pilot fallback. that disclosure is itself a form of evidence: a manufacturer willing to disclose its reliance on humans is usually doing so because the alternative is overstating capability.
Other manufacturers have been less transparent. Tesla’s We, Robot event in October 2024 featured Optimus units that, according to subsequent reporting, were teleoperated for portions of the demonstration. Tesla did not disclose this on stage. Figure has shown end-to-end autonomous behavior in its BMW pilot, but consumer-facing clips often include scenarios where teleoperation status is unclear.
Across press coverage and third-party reviews, the pattern is consistent: when a humanoid demo looks dramatically more capable than the published research base would suggest, teleoperation is the most likely explanation.
When teleoperation stops being a tell
Teleoperation is not always a problem. It is the standard data collection mechanism for behavior cloning and imitation learning, and the field expects autonomous behavior to emerge from extensive teleop datasets over time. Foundation models like RT-2, OpenVLA, and Figure’s Helix are trained substantially on teleoperated demonstrations.
The shift from teleoperated demo to autonomous behavior is gradual. A robot may be teleoperated in 2026 and run a comparable behavior autonomously in 2027 or 2028. Robovations classifies what the robot does today, not what its training pipeline projects.
The cleaner signal that a humanoid has crossed a Ladder threshold is documented end-to-end task completion in conditions the robot was not specifically prepared for. That standard is high. As of early 2026, no consumer-facing humanoid robot meets it consistently across novel households or workplaces.
What this means for buyers and observers
The Autonomy Ladder is not a verdict on whether humanoid robots are useful or whether the trajectory is real. It is a description of where they are right now. classifying teleoperation accurately is the only way the classification framework keeps its meaning. A humanoid that requires a pilot is a tool of the pilot. A humanoid that runs autonomously in scripted conditions is doing something different. A humanoid that handles novel tasks unprompted is doing something different again.
The marketing language does not reliably distinguish among these. The Ladder does. Until evidence supports a higher classification, the conservative position is also the accurate one: a teleoperated humanoid is an impressive tool. It is also not yet an autonomous robot.
