A robot can take a clear path across an office and still be a nuisance. Avoiding a chair is one problem. Passing a reading desk without a burst of mechanical footsteps is another. The destination is the same; the experience of everyone nearby is not.
A newly announced preprint, TACET, combines choices about gait, speed and route. Submitted October 2 and announced on arXiv October 6, it is a research result, not a product launch.
Quiet at the feet is not quiet at the desk
The first lever is how the foot lands. In MUTE, published as a preprint in 2025, researchers trained a quadruped controller to reduce foot-ground impacts without changing its hardware. A tunable quietness setting balances softer movement against agility. This is software changing the way a machine uses its existing legs, rather than adding insulation or a new motor.
The second lever is where the sound travels. ANAVI, a 2024 research project, predicts what a listener would hear from visual information about an indoor space, then uses that estimate to choose a route. Its real-robot demonstration used previously scanned panoramas and a teleoperated quadruped. It also proposed adjusting acoustic priorities according to what people were doing, including concentrated work.
These are different interventions. Gentler contact changes the sound source. A different path changes the listener's exposure. Neither alone tells the robot how much inconvenience a particular situation warrants avoiding. Nor are their decibel results interchangeable: a meter riding on a machine and a meter beside a person answer different questions.
The new piece is the connection
TACET links visual scene reasoning to a separate fast controller through compact gait, speed and social-spacing choices. The slower model does not issue individual motor commands. The point is to coordinate the route with the way the robot walks, rather than leave quietness at one fixed setting.
At a commanded 0.8 meters per second, mean noise was 74.4 dBA with the default Go2 controller versus 65.1 with TACET. Actual speeds were matched; each condition had ten trials. That microphone was five centimeters above the robot's body, not at a bystander's ear. Separate person-facing tests measured sound at the listener.
The gentler trip can take longer. Near a working person, TACET took 36.8 seconds versus 21.2 for the fastest comparison method. These are selected trials on one platform; scene inference used a desktop GPU. The human assessment involved 15 people rating videos, not living with the robot through a workday.
Count the interruption, not just the arrival
For someone sharing the room, the useful question is not whether a robot has one impressive quietness score. It is whether the system chooses an appropriate compromise, repeatedly, while still finishing its job.
A convincing next test would report both listener exposure and task time across ordinary working days, including mistakes in reading the situation. Quietness should not become an excuse to conceal lost throughput; speed should not make interruptions disappear from the evaluation.
That is the connection to measuring robots outside the demo: completing a task and being tolerable company deserve separate evidence.
Explore more in Robotics.
AI-assisted. Sources checked. Feature image is a conceptual AI-generated illustration, not a photograph of the experiment.
