A robot’s body can be in one place while the instructions guiding it originate somewhere else. To understand a performance, ask who chooses the task, who controls the movement, and who steps in when something goes wrong.
Teleoperation
In teleoperation, a person provides live control from a distance. The interface might translate a person’s motions or send commands through a controller. Software can still handle important parts of the movement, such as balance or the mapping between different bodies.
The H2O research paper provides a concrete example. The authors describe using an RGB camera and a learned control system for real-time whole-body humanoid teleoperation. The human supplies the motion to follow, while the robot’s controller turns that input into movement. Read the H2O paper ↗
Autonomy has a scope
For this publication, an autonomous task means the system selects and executes actions within the stated task without a person continuously steering those actions. That definition still leaves questions. Who supplied the goal? Was the environment prepared? Could the robot recover when an object moved?
“Autonomous” becomes much more informative when it is attached to a specific job and set of conditions. Picking a known item from a fixed position and completing a changing sequence of unfamiliar tasks put different demands on a system.
Assistance and supervision
A person might choose a destination, approve a grasp, reset a failed run, or take over during an exception. Those interventions deserve a description of their own. A useful report gives the kind of intervention, how often it occurred, and how much operator time it required.
Training and live operation are separate questions
A system can learn from human demonstrations and later run a task autonomously. A polished robot motion can also be controlled live. Knowing how training data was collected does not, by itself, answer who controlled a particular recorded run.
What we will label
Our coverage will use the source’s disclosed operating mode and identify the source. When control details are missing, we will mark them as undisclosed. When a demonstration mixes modes, the explanation should say which parts used each mode. Readers should be able to understand the assistance a result depended on.
Sources & reading
Sources reviewed November 5, 2026. Company claims and author-reported research are attributed in the text.