Is This Robot Autonomous or Teleoperated? A Video Evidence Guide
A practical guide for classifying robot videos using visible evidence, autonomy categories, teleoperation signals and a reusable scoring grid for current demos.
Introduction
A robot can complete the same visible task through at least five different control paths. A person may drive every joint through a VR headset. A learned policy may execute a trajectory collected from human demonstrations. A supervisor may intervene only after a failed grasp. The motion may be a fixed script triggered by a button. A vision-language-action model may close the loop from camera images to motor commands. The finished clip can look equally smooth in every case.
The useful question is therefore not whether the movement looks human. It is what information entered the system, who selected each action, when a human could intervene and whether the robot recovered from changes it did not see during training. This guide provides a repeatable classification, visible warning signs and a scoring grid. It also sets a hard limit: a short edited video rarely proves autonomy by itself. Good reporting must combine the footage with a method statement, control interface, intervention log and task conditions.
Key findings
- Smooth motion is not evidence of autonomy; direct teleoperation can be low-latency and visually natural.
- Training data collected by teleoperation can later produce autonomous execution. The collection method and execution method must be reported separately.
- Cuts, speed changes and prepared objects reduce the evidentiary value of a demo, but do not prove deception.
- A human supervisor who approves plans or recovers failures creates a human-in-the-loop or supervised-autonomous system, not pure autonomy.
- The strongest evidence is a continuous run with disclosed inputs, intervention count, resets, task duration and failures.
Robot demo classification grid
Use the narrowest category supported by the available evidence. Do not upgrade a classification because the company uses the word AI.
| Category | Who chooses the action? | Typical evidence | Correct reporting language |
|---|---|---|---|
| Autonomous | The robot policy perceives, plans and executes without human action selection during the evaluated run. | Uncut run, disclosed sensors, zero interventions, recovery behavior, task metrics. | The robot completed this specific run autonomously under the stated conditions. |
| Supervised autonomous | The robot acts independently while a human monitors and can stop, approve or redirect. | Supervisor console, approval gates, remote safety operator, intervention log. | The robot operated autonomously under human supervision. |
| Human-in-the-loop | Human input enters the task during execution, often at uncertain points or failures. | Remote assistance, corrections, escalations, handoff between policy and operator. | The task combined autonomous execution with human assistance. |
| Teleoperated | A human continuously selects motion or end-effector commands. | VR headset, motion-capture suit, gloves, joystick, mirrored movement, operator station. | A human operator controlled the robot remotely. |
| Scripted or pre-programmed | A fixed sequence or trajectory determines the motion. | Identical repetition, known coordinates, no perception-led adaptation, trigger-based routine. | The robot executed a pre-programmed routine. |
| Simulation only | A controller runs in a virtual environment, not on the physical robot shown in the claim. | Simulator interface, rendered environment, no physical hardware validation. | The policy was evaluated in simulation only. |
| Autonomy not disclosed | The publisher does not state the control mode. | Polished video without method notes or operator information. | The control method was not disclosed. |
| Insufficient evidence | Evidence conflicts or is too incomplete to classify. | Edited fragments, missing context, ambiguous overlays, no primary source. | Available footage is insufficient to determine the control mode. |
A 20-point evidence scorecard for readers
| Evidence item | 0 points | 1 point | 2 points |
|---|---|---|---|
| Continuity | Heavy edits or montage | Mostly continuous with minor cuts | Full uncut task |
| Control disclosure | No statement | General statement only | Exact interface and human role |
| Interventions | Not reported | Says none or few without count | Count, timing and reason published |
| Environment | Staged and fixed | Some variation | Multiple unfamiliar or randomized layouts |
| Error recovery | Failures removed | One recovery shown | Repeated recoveries and failure cases |
| Task metrics | No duration or success rate | Selected metrics | Trials, denominator, success and time |
| Sensor disclosure | Unknown | Sensor list only | Sensor inputs linked to policy |
| Human equipment | Off-camera area hidden | Partial workstation view | Operator area and interfaces shown |
| Speed | Unknown acceleration | Speed multiplier labeled | Real-time sequence |
| Independent access | Company video only | Partner confirmation | Third-party repeat or customer evidence |
Direct teleoperation has recognizable hardware, but it can also be hidden
The clearest teleoperation setups use a VR headset for viewpoint, hand controllers or gloves for arm and finger pose, and a motion-capture system for the upper body. Some systems map a human’s joint positions directly to the robot. Others convert the operator’s motion into end-effector targets while a controller handles balance and collision limits. The second case is still teleoperation because the human continuously chooses the task motion, even though low-level stabilization is automatic.
Look for synchronized starts, pauses and corrections between a person and robot. A delayed copy of the operator’s wrist rotation, elbow movement or grasp is strong evidence. A robot that freezes whenever the operator freezes is another clue. Joysticks and gamepads often produce more segmented translation and rotation, although experienced operators can make them look smooth.
The operator may be in another room or another country. Absence from the frame proves nothing. Product pages that mention remote experts, supervision stations, fleet operations or recovery services should be read closely because those systems may deliberately combine autonomy with human support.
Imitation learning separates demonstration from execution
Teleoperation is commonly used to collect training data. The operator opens a drawer, folds a towel or loads a dishwasher while cameras, joint positions and forces are recorded. A policy is then trained to predict actions from observations. When the trained policy later controls the robot without live human commands, that evaluation run can be autonomous even though every training example came from teleoperation.
Google DeepMind’s ALOHA work follows this pattern: human demonstrations teach a bimanual platform, after which learned policies execute tasks. Physical Intelligence similarly trains generalist policies from robot data and evaluates them through autonomous rollouts. The correct caption should say both things: demonstrations were collected through teleoperation, and the published evaluation was autonomous under the stated setup.
Data collection footage is often mislabeled as a capability demo. Gloves, leader arms or motion capture may show how examples are generated, not how the final policy runs. Check the video title, paper methods and timeline before classifying it.
Shared autonomy and human-in-the-loop systems divide the task
Shared autonomy means the human supplies intent while the robot handles part of the execution. An operator may point to an object and let the robot plan the grasp. A person may drive the mobile base while autonomous collision avoidance controls local motion. A language command may select a goal while a manipulation policy decides joint actions.
Human-in-the-loop is broader. It includes approval before a risky action, corrections during a task, escalation after a confidence threshold and remote recovery from failure. This architecture can be commercially sensible because it keeps a deployment running while autonomous coverage improves. It should not be described as autonomous without the qualifier.
Supervised autonomous is appropriate when a robot normally acts on its own but a human continuously monitors safety or operations. The difference from human-in-the-loop is whether human input actually changes the evaluated run. Publishers should report both the availability of intervention and the number of interventions used.
Scripted movement, reinforcement learning and VLA policies solve different problems
Scripted motion
A pre-programmed trajectory can be impressive and technically demanding, especially for dynamic locomotion. It does not require the robot to identify a changed object or choose a new plan. Exact repetition, fixed foot locations and failure after a small scene change are common signs.
Reinforcement-learning policy
A reinforcement-learning controller may learn balance, locomotion or manipulation through rewards in simulation or on hardware. During deployment the policy can be autonomous, but a joystick may still provide desired velocity or heading. Report the learned layer and the human command layer separately.
Vision-language-action policy
A VLA model consumes images and a language instruction, then outputs robot actions or action chunks. That architecture supports closed-loop autonomy, but the label VLA does not prove the shown run lacked operator corrections, safety approvals or resets.
World model and planner
A world model predicts possible future states. A planner may use those predictions to choose actions, while a separate low-level controller executes them. Some projects call the complete stack a model; others disclose each component. Classification should follow the actual action-selection path.
Visible clues that lower confidence in an autonomy claim
Frequent cuts can remove failures, resets and human repositioning. Speed-up can conceal a 20-minute manipulation sequence or make pauses disappear. Neither practice is automatically deceptive when clearly labeled, but both prevent timing and continuity claims.
Prepared environments reduce variation. Objects may be placed inside a narrow grasp region, drawers may be left partially open and visual markers may simplify localization. Repeating the same task with the same object pose does not test generalization. A stronger demo randomizes the scene and reports how many trials failed.
Rigid trajectories, no reaction to a displaced object and identical timing across repetitions suggest a script. A robot that never attempts recovery may have had failed trials edited out, or its controller may simply stop on uncertainty. Commands spoken in a video do not prove speech caused the action; the audio may be a trigger, a later overlay or a genuine language input. The method statement must decide.
Limitations and missing information
- A video cannot reveal an operator who is outside the frame or connected remotely.
- Latency varies with network, controller and filtering, so visible delay is not a reliable standalone test.
- A learned policy can repeat trajectories consistently; repetition alone does not prove scripting.
- An uncut video can still use prepared object poses, hidden initialization and human approval before recording starts.
- “Autonomous” can describe navigation while manipulation remains teleoperated, so each subsystem and task phase must be classified.
Conclusion
The safest classification follows the action path, not the marketing label. Identify the sensor inputs, find who selected the goal, determine who selected the motion and record whether a human intervened. Teleoperation means the person continuously drives the action. Human-in-the-loop means human decisions enter during execution. Supervised autonomy means the robot acts while a person monitors. Scripted motion follows a fixed routine. Autonomous execution requires the robot to close the loop through perception and action for the evaluated run.
When those facts are missing, “autonomy not disclosed” is more accurate than guessing. A transparent company should be able to state the run length, edits, speed, number of trials, resets, interventions, data-collection method and evaluation environment. That information turns a compelling clip into usable technical evidence.
Frequently asked questions
How can I tell if a humanoid robot is teleoperated?
Look for VR headsets, motion-capture suits, gloves, leader arms, joysticks and movements synchronized with a person. Also search the official project page for words such as teleoperation, expert, remote supervision or data collection. None of these clues alone proves the published task was teleoperated, because the same equipment may have been used only for training data.
Does an uncut video prove autonomy?
No. An uncut video is stronger than a montage because it preserves duration, pauses and visible failures. It still may begin after human setup, use a prepared environment or receive remote commands. Control disclosure and intervention reporting are required before the run can be classified confidently.
Is imitation learning the same as teleoperation?
No. Teleoperation is a control method in which a human drives the robot. Imitation learning is a training method that learns from demonstrations, often collected through teleoperation. A trained imitation policy can later execute a task autonomously, with human supervision or with live assistance depending on the deployment.
What is shared autonomy in robotics?
Shared autonomy divides control between a human and the robot. A person may choose the object, goal or direction while the robot plans a collision-free path, stabilizes the body or executes the grasp. Reporting should specify exactly which decisions came from the person and which came from the autonomous controller.
Can a robot be autonomous if a safety operator is present?
Yes, for a specific run, provided the safety operator does not select actions or intervene. The precise label is often supervised autonomous because a human is monitoring and can stop the system. The publication should report whether any intervention occurred and what would trigger one.
What evidence is strongest for an autonomy claim?
The strongest package combines continuous real-time video, randomized task conditions, disclosed sensor and policy inputs, trial counts, success rate, duration, intervention and reset logs, failure examples and a clear separation between training data collection and evaluation. Independent repetition by a customer or laboratory adds further confidence.
Sources and methodology
The guide uses a control-responsibility test: who chooses the goal, who chooses each action and when a human can change the run. Definitions were checked against official robotics demonstrations and model documentation. The scorecard measures disclosure quality, not the underlying robot’s intelligence or commercial value.
- NEO product and Expert Mode — 1X Technologies · Accessed July 11, 2026
- Helix 02 full-body autonomy — Figure AI · January 27, 2026
- Advances in robot dexterity — Google DeepMind · September 12, 2024
- Open X-Embodiment repository — Google DeepMind · Accessed July 11, 2026
- π0 generalist policy — Physical Intelligence · October 31, 2024
- OpenVLA repository — OpenVLA team · Accessed July 11, 2026
- RT-2 vision-language-action model — Google DeepMind · July 28, 2023
Related TechniaHQ guides
Official image recommendations
- Split-screen showing a teleoperation station and a robot executing alone.
Robot operator wearing a VR headset beside a separate autonomous robot demonstration — Official manufacturer and laboratory media - ALOHA demonstration collection hardware.
Human operator using a bimanual leader setup to demonstrate a robot task — Google DeepMind - NEO Expert Mode disclosure and home robot view.
1X NEO in a home with the official Expert Mode description — 1X Technologies - Action-responsibility flowchart from human command to motor output.
Decision tree for classifying autonomous, supervised, human-in-the-loop and teleoperated robots — TechniaHQ original graphic - 20-point transparency scorecard formatted for mobile.
Robot demo evidence scorecard with ten two-point criteria — TechniaHQ original table
Fact-check report
Verified: July 11, 2026
Confirmed
- The classification definitions distinguish action selection from training-data collection.
- Examples link to primary project pages rather than social reposts.
- The article does not classify an unidentified video without supporting method information.
Not confirmed or incomplete
- A viewer cannot detect all remote intervention from pixels alone.
- Companies use the word autonomous with different subsystem boundaries and evaluation conditions.
Fast-changing information
- Robot fleet supervision tools and intervention policies change as software is updated.