UCLA's PhoneBot Puts a Humanoid Robot in a University Club's Budget

A bipedal humanoid robot that walks, recovers from a fall, follows a person with its camera, and answers spoken commands can now be built for under $400 in hardware. The trick is not a cheaper motor or a simpler leg. It is the phone in your desk drawer. Researchers at UCLA's Robotics and Mechanisms Laboratory have open-sourced a platform called PhoneBot that uses a commodity Android smartphone as its sensor array, processor, radio, and display, all at once.

The cost gap PhoneBot represents is the story. The closest comparable open-source bipedal research platform costs $3,500, and the most widely used academic competition humanoid runs to $13,800, a 9x to 35x reduction that moves this class of hardware from the funded-lab budget into the university club budget.

Five subsystems, one prepaid handset

A small research humanoid normally needs five or more separate subsystems: a depth camera or stereo rig for vision, an inertial measurement unit for balance sensing, a single-board computer for inference, a wireless radio for remote operation, and a screen or speaker for interaction. Each part adds cost, wiring, and failure points.

PhoneBot's answer is architectural. The phone's rear camera handles vision. Its built-in IMU, the same chip that tracks screen orientation and motion in fitness apps, provides the real-time orientation and angular velocity data that stable bipedal walking requires. The phone's application processor runs the neural-network locomotion controller and on-device vision models at 50 Hz. Its Wi-Fi radio handles all communication, and its touchscreen and speaker handle interaction. One piece of consumer hardware, priced for the prepaid market, replaces a subsystem bundle that would cost hundreds of dollars even at academic pricing.

The machine itself is small: 483 millimeters tall (about 19 inches) and 1.8 kilograms including phone and battery, roughly the size and weight of a large house cat. It carries 13 servo motors, six joints in each leg plus a single torso-yaw joint that lets the phone's camera track a person without pivoting the whole robot. The servos are Dynamixel XL430-W250-T units, a standard academic robotics component available off the shelf.

One detail shows real engineering judgment. The phone does not connect directly to the motors. It talks over a local Wi-Fi link to a small single-board computer in the lower body, which translates commands into the serial signals the servo bus needs. That wireless indirection means the phone is swappable without touching the robot's electronics, treating a component that will be upgraded or replaced as an interchangeable peripheral rather than a permanent part of the stack.

One piece of consumer hardware, priced for the prepaid market, replaces a subsystem bundle that would cost hundreds of dollars even at academic pricing.

How it learns to walk

Bipedal locomotion cannot be reliably hand-coded. The physics of an inverted pendulum walking on two legs is too sensitive to small perturbations and hardware tolerances. PhoneBot's controller was instead trained with reinforcement learning in the MuJoCo/MJX physics engine: a virtual robot earned reward for maintaining forward speed, alternating steps correctly, and staying upright, and lost reward for dragging its feet, consuming excess energy, or falling.

The algorithm is an asymmetric actor-critic variant of Proximal Policy Optimization, where the critic, used only during training, sees privileged simulator information such as exact foot-contact states that the deployed policy never gets. The simulation was parameterized to match the real hardware's limits, servo torque ceilings, speed limits, gear backlash, and those values were randomized during training so the resulting policy would be robust to real-hardware variation. The final policy runs on the phone via TensorFlow Lite at 50 Hz, issuing fresh joint-angle targets every 20 milliseconds.

The crooked-walk fix

The paper's most transferable contribution solves a problem every builder of a cheap biped eventually hits: the real robot walks crooked even when the simulation robot walks straight. The cause is manufacturing tolerance. Two servos rated to identical specifications have slightly different stall torques, gear ratios, and friction, and a policy trained on a perfectly symmetric virtual robot inherits none of that asymmetry.

The team's solution, called sagittal mirroring, fixes this during training rather than in hardware calibration. After each training step, the policy's parameters are averaged with a left-right mirrored version of themselves, forcing the policy to behave well in both the original and the mirror image of the robot. The numbers are striking: joint-angle symmetry error fell from 0.254 radians to 0.003 radians, a 98.8 percent reduction, and left-right torque imbalance dropped 83.4 percent. The technique applies to any low-cost biped trained in symmetric simulation, which is virtually all of them.

Any phone, almost

The team tested four Android devices: a moto g 2025 (about $50), a Samsung Galaxy A16 5G ($50 to $80), a moto g 2024 (about $50), and a 2017 Honor 9, included to probe the limits of old hardware. All four handled the core functions: balance sensing, locomotion control, on-device human detection, and speech interaction. The Honor 9's only failure was the SLAM mapping feature, which needs Google's ARCore framework; the Honor 9 cannot access Google services in the US, a distribution restriction rather than a processing limitation.

The lineage here is real. PhoneBot extends the idea Intel Labs demonstrated with OpenBot in 2020, a wheeled robot controlled by an Android phone. A wheeled robot is statically stable; a biped must actively manage dynamic balance through every step. Demonstrating that phone-class hardware can satisfy those real-time constraints, including fall recovery, is the substantive extension.

What it cannot do, and why that matters less than it sounds

The paper is candid about three limitations. The robot has no arms, which removes angular-momentum compensation that helps bipeds turn and recover from pushes. The walking policy was trained and validated only on flat terrain. And the conversational AI currently routes language-model inference through an external laptop over Wi-Fi rather than running on the phone, though on-device LLM inference is an explicit near-term goal.

Those gaps define a research platform, not a product, which is the point. RoMeLa's open-source tradition runs deep: DARwIn-OP became a standard RoboCup platform, and the lab's full-size ARTEMIS humanoid won the Adult-Size RoboCup championship in 2024, making RoMeLa a six-time world champion. PhoneBot targets a different point in the cost-capability spectrum, one designed for maximum accessibility.

The larger lesson is the one the eeDesignIt coverage put best: engineers now have a mature, tightly integrated module for perception, compute, communications, and interaction sitting in a drawer. PhoneBot asks how much of a robot still needs to be designed from scratch. The full hardware designs, firmware, and software are available at phonebot.dev, so any student with a soldering iron and a spare prepaid phone can go find out.