PhoneBot paper puts an Android phone in charge of a biped
UCLA researchers presented PhoneBot in an October 6 2026 preprint. The biped has 13 actuated joints and an estimated $400 hardware cost excluding its smartphone.
Its Android app supplies sensing, vision and locomotion inference. It sends motor targets over WiFi to a separate onboard relay computer. Dialogue uses a laptop backend. The policy issues joint targets at 50 Hz.
How the walking evidence works
The training method models servo torque limits and adds mirrored leg movements to learning samples. Across five training seeds on flat terrain, the authors report improved consistency against otherwise equivalent training without mirroring. These comparisons are simulation results.
The paper also shows physical walking, standing and person tracking, and reports application tests on four Android phones. PhoneBot has no arms, while limited motor torque constrains payload and dynamic movement. ARCore localization was unavailable on the Honor 9 test device without Google services.
What motor ratings tell builders
The ROBOTIS actuator manual helps explain why motor specifications need careful reading. It distinguishes stall torque from continuous output. Stall torque is a momentary maximum, and the company says actual performance is generally closer to the measurements in its performance graph. A headline torque rating therefore cannot establish how much weight a completed robot will carry while walking.
The manual also says its reported load value is inferred from internal output rather than measured by a torque sensor. That distinction matters when assessing force measurements. Monitoring a servo estimate is different from independently measuring contact force or payload. A reproduction should keep those quantities separate when reporting results.
A phone was already a useful robot computer
The earlier OpenBot paper, first submitted in August 2020, paired an Android phone with a small wheeled vehicle. Its authors demonstrated person following and autonomous navigation. This provides useful historical context for smartphone reuse without implying that locomotion results transfer between wheels and legs.
For readers considering a first build, the OpenBot repository offers a concrete reference for the work involved. Its setup path covers constructing the body, flashing Arduino firmware, installing Android applications and training a driving policy. Those are separate integration tasks. An existing phone reduces the need to acquire some electronics, but it does not remove assembly, software installation or testing.
Check the build materials before ordering parts
On October 13, the readable PhoneBot project page exposed a hardware figure without accessible code or build downloads. We could not independently verify a complete public release of those materials. Before buying components, confirm that the actual design files, software, instructions and their licenses are available. This check does not establish that the files are absent elsewhere.
Related ByteForward coverage of robot skill training through simulated practice also distinguishes simulator evaluation from physical trials. That distinction is useful when reading robotics results.
Featured photograph of an Android smartphone by Jonas Leupe on Unsplash, used under the Unsplash License. Illustrative image rather than PhoneBot hardware.
