A 19‑inch humanoid walks and sees with a phone

PhoneBot works by handing nearly every job to the phone mounted in its torso. The camera handles vision, so the robot can spot and follow a person. The phone’s inertial measurement unit, a chip that senses tilt and motion, acts like an inner ear for balance. Its processor runs the neural network (a type of AI model) that decides how the legs should move, and the touchscreen shows status and controls.

The phone isn’t plugged into the body. It sends commands over Wi-Fi to a small onboard computer, which drives 13 low-cost servo motors: six in each leg and one that swivels the torso. That swivel lets the camera turn toward a person without the robot shuffling its feet around. Using Wi-Fi instead of a cable also makes it easier to swap in a different phone.

The robot stands about 19 inches tall and weighs around four pounds with the phone and battery included, small enough to sit on a desk. It was developed by UCLA’s Robotics and Mechanisms Laboratory, known as RoMeLa, and described in a research preprint.

You don’t need a flagship phone. The team tested four Android devices: an Honor 9 from 2017, a Moto G 2024, a Moto G 2025 and a Samsung Galaxy A16 5G. The newer three are prepaid models that cost about $50 to $80. All four handled the core jobs: sensing motion, controlling the walk, running vision on the phone and interacting with people.

There was one exception. Camera-based mapping, which lets the robot build a picture of a room and track where it is, didn’t run on the Honor 9 because the Google services it needs weren’t available on that phone in the US. The researchers don’t claim PhoneBot will run on every Android device, so it’s worth checking a phone’s hardware and software before you commit.

Two legs are much harder than wheels. Earlier phone-powered robots, such as Intel’s wheeled OpenBot, didn’t have to worry about falling over. PhoneBot did.

The team first trained it in simulation using reinforcement learning, a trial-and-error method where software earns points for good behavior and loses them for bad habits. PhoneBot scored for keeping pace, alternating steps and staying upright, and lost points for things like dragging its feet. The simulation included the real motors’ limited strength, speed and gear slack, so the robot only learned moves its body could actually perform.

Even with a symmetrical design, the first gait came out uneven, with one leg doing more work. The fix was simple: mirror the training examples so both legs learned the same patterns. That balanced the load without having to retrain from scratch. On the real robot, the team showed it walking, following a person with its camera and getting up after falling face down.

PhoneBot isn’t going to fetch your coffee. It has no arms, its walking only works on flat ground, and its cheap motors limit how much weight it can carry and what moves it can make.

Conversation isn’t fully on the phone yet either. Speech recognition and the language model run on a nearby laptop that the phone reaches over Wi-Fi. Walking and vision already run on the handset.

The researchers plan to tackle uneven terrain, add lightweight arms and use language models to turn spoken instructions and what the camera sees into actions. They also want to run suitable AI models directly on the phone so the laptop isn’t needed.

The price is the headline. Without the phone, PhoneBot costs roughly $400 to build. Other small research humanoids listed in the team’s paper range from about $3,500 to $13,800. They can do different things, so it isn’t a perfect comparison, but it puts a walking robot within reach of a classroom budget.

The hardware and software are open source, which means anyone can download the designs, build one and modify it. For schools and hobbyists, that’s a practical way to learn robotics and AI on real hardware. For everyone else, it’s a reminder that a phone too old for daily use still has serious computing power. If you have one in a drawer, it may be worth keeping rather than tossing.