Old Android phones gathering dust in drawers could soon have a second career as the brains of a walking robot, thanks to an open-source project from UCLA.
The robot, called PhoneBot, stands about 19 inches tall and weighs less than 4 pounds. Researchers at UCLA’s Robotics and Mechanisms Laboratory mounted a commodity smartphone as its torso and used the phone’s camera, motion sensors, processor and wireless connection for vision, balance sensing and movement calculations.
$400 body for any phone
The body, motors and supporting electronics cost around $400, leaving the phone as the main variable. The electronics are kept simple on purpose, so swapping one phone for another requires no rewiring.
PhoneBot uses 13 low-cost DYNAMIXEL servo motors: six in each leg, with joints at the hip, knee and ankle, plus one that spins the upper body. That torso joint lets the phone’s camera sweep from side to side while the feet stay planted, which helps when tracking a person. A small computer in the lower body receives commands from the phone over Wi-Fi and converts them into motor positions, so the phone never has to be wired to the hardware.
Four phones, four passes
The team tested four Android devices, including a 2017 Honor 9 and two inexpensive off-the-shelf models priced between $50 and $80. All four handled the core jobs of walking control, motion detection and basic vision. The older Honor lacked the Google services needed for onboard mapping.
Taught to walk in simulation
The walking controller is a neural network trained entirely in simulation with reinforcement learning. The virtual robot earned rewards for staying upright, alternating feet and holding speed, and lost points for high torque, harsh braking and loading one foot. Real motors have limited strength and speed, so those constraints were built into the simulation.
Early results were uneven, with one foot doing most of the work. Adding mirrored samples that swapped left and right produced a more even stride, less variation between runs and better load sharing. The physical robot can walk in any direction on level floors at up to half a meter per second (about 1.6 feet per second) and push itself upright after a fall.
Everything runs on the phone
An Android app makes decisions in real time. TensorFlow Lite runs the walking and recovery routines about 50 times a second using the phone’s built-in sensors. The camera tracks people and their positions, and on supported phones it can also perform visual mapping for telepresence and photography.
Speech recognition and language processing still run on a nearby laptop, though the team could move them onto the phone as models improve.
What it cannot do yet
PhoneBot has no arms, so it cannot pick up or carry objects. It has been tested only on flat ground, and stairs, slopes and uneven surfaces remain works in progress. Limited motor power also caps the weight it can carry and how fast it can move.
The hardware and software are fully open, available to students, research labs and hobbyists who want to work with bipedal robots without paying for pricey specialized gear or sensors.