Humanoid robot guide
Are humanoid robots teleoperated or autonomous?
Some humanoid systems support both autonomous and remote control. Evaluate the particular task and demonstration, including what the human does, instead of assigning a single label to the whole robot. [s3]
By humanoidversus Editorial Desk
Published · Last verified Oct 13, 2026
TL;DR
- A robot can support multiple control modes: Boston Dynamics lists autonomous, teleoperated and tablet steering modes for its product Atlas. [s3]
- Teleoperated training data does not mean an evaluation is remotely controlled. HumanPlus separates human shadowing from learned autonomous skills. [s7]
- 1X describes scheduled Expert assistance for unfamiliar NEO chores; its order page also advertises user remote control. [s1] [s2]
- Check control, editing and timing disclosures. Boston Dynamics describes an uncut policy demonstration and separately explains faster physical policy execution. [s5]
Autonomous, teleoperated and scripted mean different things
Some humanoid robots are teleoperated for certain tasks and autonomous for others. In its January 2026 product announcement, Boston Dynamics lists autonomous control, teleoperation and tablet steering for Atlas. Ask how the particular demonstration was controlled; its robot name does not supply a permanent autonomy classification. [s3]
Use teleoperated for actions directed by a human during execution, such as tracked hand movement or joystick commands. Unitree's XR teleoperation project documents these inputs and distinguishes arm control from walking controls. Software can still handle part of the motion. Ask which parts the person directs and which the robot controls, because a remote-controlled task need not mean every joint is manually commanded. [s8]
Use autonomous task execution for the scope the robot completes without ongoing human action commands. A person can still provide a goal: the Atlas engine-cover demonstration takes bin-location inputs while its maker says movement is generated online without prescribed or teleoperated motions. Use scripted for a predefined action sequence in your comparison notes, and ask how much sensing or adaptation remains inside that sequence. [s4]
Separate learning from control during the demonstration
A human demonstration can be a training example rather than the source of live commands in a later test. Stanford's HumanPlus project explicitly separates shadowing human motion to collect data from training policies that perform skills autonomously. When a paper or video mentions teleoperation, first identify whether it describes data collection, the demonstrated evaluation, or a fallback during execution. Those are different claims. [s7]
Figure's Helix announcement describes a teleoperated dataset and learned control using images, language and robot state. Training descriptions alone do not establish how every clip was produced. Check each caption and evaluation setup, and attribute the claimed control method to that demonstration. A model name should not substitute for independent verification of autonomy. [s6]
Keep the scope of an autonomy claim narrow. In its original Helix announcement, Figure describes upper-body control; HumanPlus presents specific learned skills. Ask separately about movement to the work area, object selection, manipulation, recovery and the next task. Our glossary links for embodied AI and vision-language-action models explain the terminology, but those categories should not replace evidence about the actual behavior being evaluated. [s6] [s7]
What official disclosure looks like
A useful disclosure names the input, the controlled activity and the human's role. Boston Dynamics' October 2024 Atlas Goes Hands On description identifies the engine-cover task, the bin-location input and sensing used to respond to changes and failures. The maker says the robot chooses its movements during execution without a preset motion sequence or live operator control. This predates the 2026 product announcement and remains a task-specific official claim rather than an independently reproduced result. [s4] [s3]
Another useful format distinguishes control modes before the demonstration. Unitree's public XR project identifies teleoperation in its title and introduction, lists supported arm configurations and explains tracked-hand and controller inputs. That documentation supports the existence and structure of the remote-control system. It does not establish that every Unitree promotional video uses that system, so avoid extending the disclosure to clips it does not identify. [s8]
Research explanations can also identify the transition from demonstration to learned behavior. HumanPlus separates its autonomous-skills and human-shadowing material. When summarizing a mixed collection of clips, retain those labels rather than describing the entire project as either remotely controlled or autonomous. For a buyer, request similarly explicit labels alongside the demonstration of the exact configuration and software offered. [s7]
NEO Expert Mode and human assistance at home
1X's NEO launch description says owners can schedule an Expert to guide the robot through unfamiliar chores, alongside basic initial task abilities that are intended to improve over time. Treat assisted chores and independently performed chores as separate parts of the offer. Before relying on a household demonstration, ask which mode produced it and what assistance the delivered product would need in your home. [s1]
The NEO order page describes scheduled remote expert supervision and separately advertises piloting the robot through a mobile app and VR device. A remote assistance feature does not establish the control method of every NEO video. Request a task-by-task explanation of who provides commands, when a person can join, and what happens if a task cannot be completed without assistance. [s2]
For an assisted product, request its consent and privacy workflow: who can access camera or microphone feeds, how sessions start and stop, and what data is retained. Resolve those questions in service documentation rather than inferring assurances from Expert Mode's name. Compare household usefulness with the proposed human access and supervision arrangements. [s1] [s2]
Cuts, playback speed and physical task speed
Ask for continuous footage from instruction to completion, including setup, waiting and recovery. Boston Dynamics' August 2025 research article describes an uncut Atlas demonstration with subtasks triggered by high-level language prompts. Spot Workshop names the task; Atlas performs it. Seek the same editing and human-input disclosures for other clips. [s5]
Ask two separate timing questions: was playback accelerated, and did the robot physically execute faster? Boston Dynamics' learning article discusses changing the timing of predicted actions to increase policy execution speed. That is different from speeding up a recording. When a caption says faster or shows a multiplier, ask what was changed before using the clip to estimate physical task duration or throughput. [s5]
For operational comparisons, request elapsed time, attempted runs, completed runs and interventions under stated conditions. A selected success should be recorded as a selected success unless wider results are supplied. Figure explains exclusion of training objects from its Helix evaluations, while HumanPlus describes task-specific evaluation outcomes. Look for those accompanying methods and results rather than deriving a reliability percentage from a promotional montage. [s6] [s7]
A practical checklist for reading a robot demo
First identify the original uploader, publication date, robot revision, hardware options and software version. Then write the claimed task in plain language. Record whether the maker identifies autonomous execution, teleoperation, replayed motion or a combination. Atlas' mode disclosure and Unitree's explicit teleoperation documentation show why the platform and the control method need separate entries in your notes. [s3] [s8]
Next record the human inputs: initial instruction, movement commands, grasp commands, prompts between subtasks and interventions. Ask for a continuous view, declared playback rate, elapsed time, failures and recovery attempts. Use the Atlas descriptions as examples of useful disclosure, including specified task inputs and language-triggered subtasks. Leave undisclosed details marked unknown rather than supplying a control method based on how smooth or humanlike the movement looks. [s4] [s5]
Finally identify whether the behavior is a research evaluation, product feature or contracted task. HumanPlus's learned skills and NEO's expert-guided chores illustrate different contexts. Request a repeat demonstration with your objects and conditions when procurement depends on it, and record preparation and assistance. Keep your conclusion within the disclosed setup and results. [s7] [s1]
Quick answers
Frequently asked questions
Are all humanoid robots remote controlled?
No single control label covers every system or task. Boston Dynamics lists both autonomous and teleoperated modes for product Atlas; verify the mode used in each demonstration. [s3]
Does teleoperation mean the robot has no AI?
Teleoperation can coexist with learned control. HumanPlus uses human shadowing to collect demonstrations and then learns autonomous skills from that data. [s7]
How can you tell if a robot demo is autonomous?
Look for an explicit task-specific statement and disclosed inputs. Atlas Goes Hands On describes its bin-location input and states that its movements are generated online without teleoperation. [s4]
Is 1X NEO fully autonomous?
1X describes basic initial autonomy and scheduled Expert guidance for unfamiliar tasks. Ask which chores the delivered offer performs independently and which require remote assistance. [s1] [s2]
Can a human give instructions to an autonomous robot?
Yes. Boston Dynamics describes a learned policy demonstration whose subtasks are triggered by high-level language prompts. Record that prompting separately from live motion commands. [s5]
Does training on teleoperated data make a robot demo teleoperated?
No. Training data and execution control are separate. HumanPlus describes collecting demonstrations through shadowing and learning policies for autonomous task performance. [s7]
Why should you check whether a robot video is sped up?
Ask whether the multiplier changes playback or actual execution. Boston Dynamics describes changing policy action timing to increase physical execution speed, which is a different claim from accelerated footage. [s5]
Evidence
Sources
- [s1]NEO launch: basic autonomy and scheduled Expert-guided chores ↗1X Technologies · official · published Oct 28, 2025 · accessed Oct 13, 2026
- [s2]NEO Expert Mode and user remote-control features ↗1X Technologies · official · accessed Oct 13, 2026
- [s3]Product Atlas announcement: autonomous, teleoperated and tablet modes ↗Boston Dynamics · official · published Jan 5, 2026 · accessed Oct 13, 2026
- [s4]Atlas Goes Hands On: task inputs and control disclosure ↗Boston Dynamics · official · published Oct 30, 2024 · accessed Oct 13, 2026
- [s5]Large Behavior Models and Atlas Find New Footing ↗Boston Dynamics · official · published Aug 20, 2025 · accessed Oct 13, 2026
- [s6]Helix: teleoperated training data and learned humanoid control ↗Figure AI · official · published Feb 20, 2025 · accessed Oct 13, 2026
- [s7]HumanPlus: human shadowing, training and autonomous skills ↗Stanford HumanPlus research team · official · accessed Oct 13, 2026
- [s8]Unitree XR teleoperation: supported inputs and control modes ↗Unitree Robotics · official · accessed Oct 13, 2026
AI-assisted, editor-reviewed · Read our editorial policy and verification process.