UCLA's PhoneBot Is a $400 Humanoid That Runs on an Old Android Phone's Camera, Sensors and Chip

Researchers at the University of California, Los Angeles have unveiled PhoneBot, an open-source humanoid robot that gets its brain, eyes and balance system from an ordinary Android smartphone bolted into its torso. The point is cost: the robotic hardware, excluding the phone, comes to roughly 400 dollars, which puts a walking biped within reach of a student budget rather than a lab grant.

The robot stands 483 millimeters tall, weighs 1.8 kilograms and is driven by 13 servos, six per leg covering hips, knees and ankles, plus one joint that rotates the torso and the mounted phone. That rotation matters: the camera can turn to look at a person without the robot pivoting its whole body, which the team uses for tracking and other vision-based interaction.

The phone does the work dedicated robots usually reserve for specialized parts. Its camera handles person detection and pose estimation, its inertial measurement unit supplies the motion data needed for balance, and an Android app runs locomotion policies through TensorFlow Lite. A separate onboard computer receives commands over Wi-Fi and translates them into joint movements, a deliberate split that keeps construction and replacement simple: the smartphone is the only expensive, variable component.

The team tested four Android phones, including an Honor 9 released in 2017 and three inexpensive handsets costing about 50 to 80 dollars. All four supported the core functions of motion sensing, walking control, visual detection and human interaction, with some features varying by model. Old phones, in other words, are already fast enough to be robot brains.

Walking was learned in simulation with reinforcement learning that rewarded stable locomotion and penalized unstable posture and excessive motor load. Early training produced an uneven gait that favored one leg, so the researchers introduced mirrored training samples, teaching the controller through corresponding left and right movements. Across five independent training runs, mirrored learning cut joint-angle symmetry error from 0.254 to 0.003 radians, reduced left-right torque imbalance from 0.205 to 0.034, and brought mechanical-power imbalance down from 0.339 to 0.040.

Physical demonstrations included walking, recovering after being knocked face-down, and aiming the phone camera at people without moving the feet. The limits are honest and stated: low-power servos restrict payload and dynamic movement, the current controller assumes flat ground, and the platform has no arms for manipulation. The research is available as a preprint on arXiv and has not yet been through peer review.

PhoneBot is not competing with Tesla's Optimus or Agility's Digit. Its significance is the floor it sets: a functioning bipedal research platform for the price of a decent monitor, built from hardware most people already own and are about to discard. For teaching humanoid locomotion and letting labs iterate on control software without six-figure hardware budgets, that floor may matter more than any benchmark.