Are humanoid robots autonomous?

Separate the decisions

A robot can keep its balance automatically while a person chooses where it walks. It can also receive a destination from a person and plan the movement itself. The meaningful question is which decisions the machine makes during a specified job. The following breakdown is a reading aid, not an official certification scale.

Atlas has a documented task boundary

In Atlas Goes Hands On, Boston Dynamics describes the demonstrated robot moving engine covers between supplier containers and a sequencing dolly. The input is a list of bin locations. The company says a vision model locates the fixtures and bins, a grasping policy handles the covers, and motions are generated online without teleoperation. It describes responses to displaced fixtures, failed insertions and other contact problems. [1]

That account identifies a task, an input and several recovery behaviours. It does not establish that Atlas can inspect any workshop, decide what should be manufactured and run the whole process. For a purchasing trial, the relevant extension is another cover, bin arrangement or workflow under documented conditions. Changing the task requires new evidence.

NEO combines several operating modes

1X describes NEO's early-owner capability as basic autonomy, with scheduled remote experts for unfamiliar work. That combination can contain automatic body control and human task guidance at the same time. Record the mode used for each attempt, including the moment a remote operator takes over. [2]

Training and execution also need separate descriptions. The June 2025 Redwood account says its training data included teleoperated and autonomous episodes, while the model itself ran on the robot's embedded GPU. Learning from human-controlled examples does not mean a human controls every later movement. Conversely, an onboard model does not prove that the whole task runs without help. [3]

Cloud computation is another question

The 1X world-model report describes remote inference producing a video plan and an inverse-dynamics model extracting actions. It reports 11 seconds to generate five seconds of video, followed by one second to extract the actions. Those research timings concern computation and latency. They do not describe a person steering the robot. [4]

Keep three columns in an evaluation log. First record the source of the task instruction. Next record where software executes. Finally record each person's intervention. Putting all three under the word autonomous makes systems with very different dependencies look equivalent.

Ask for an intervention log

• Name the hardware and software version, objects and permitted working area.
• Count all attempted tasks, including abandoned attempts.
• Distinguish a new instruction from manual guidance, reset or physical rescue.
• Report completion time alongside the success fraction.
• Describe behaviour during loss of communication and during an unsuccessful grasp.

A precise claim has a boundary. Moving engine covers from named bins is a boundary a reader can inspect. Working anywhere, with anything, is not established by the evidence above.