PhoneBot: An Open Source Humanoid Robot Powered by Android Smartphones
UCLA researchers leverage repurposed smartphone hardware to create a cost-effective, compact humanoid robotics platform.
1Repurposing Mobile Hardware
UCLA researchers developed PhoneBot, a humanoid robot that utilizes inexpensive Android smartphones for vision, sensing, and control. By replacing dedicated cameras and onboard computers with existing mobile hardware, the team significantly lowered entry barriers. The platform costs approximately $400 in robotic hardware, excluding the phones, making it highly accessible for makers.
2Technical Specifications Overview
Standing 483 mm tall and weighing 1.8 kg, the robot features 13 servo motors for movement. It utilizes six actuated joints per leg, covering hips, knees, and ankles, plus a torso rotation joint. This configuration allows the camera to track people independently without requiring the entire robot body to rotate during interaction.
3Smartphone Integration Logic
The system relies on the smartphone's built-in inertial measurement unit for balance and motion data. A separate internal computer manages motor communications, receiving wireless commands via Wi-Fi from the phone. This architecture translates high-level visual data into precise joint movements, effectively turning a standard mobile device into the robot's central processing brain.
4Device Compatibility Testing
Researchers successfully tested four Android handsets, including a 2017 Honor 9 and devices priced between $50 and $80. All tested models supported core functions like person detection, pose estimation, and walking control. The application runs locomotion policies using TensorFlow Lite, proving that older, low-cost hardware remains viable for complex robotics tasks.
5Reinforcement Learning Training
The team trained the walking controller using reinforcement learning in simulation, rewarding stable movement while penalizing excessive motor loading. To address initial gait asymmetry, they introduced mirrored training samples. This method taught the controller corresponding left and right movements, significantly improving symmetry without requiring additional simulated physics steps during the training process.
6Quantifiable Performance Gains
Mirrored learning yielded measurable improvements across five training runs. Joint-angle symmetry error dropped from 0.254 to 0.003 radians, while left-right torque imbalance fell from 0.205 to 0.034. Additionally, mechanical-power imbalance declined from 0.339 to 0.040, demonstrating that targeted training techniques can drastically enhance the stability of low-cost robotic platforms during physical execution.
7Current Hardware Limitations
PhoneBot faces constraints due to its low-powered servos, which limit payload capacity and dynamic movement reliability. The current controller is optimized for flat ground, and the lack of arms prevents object manipulation. Furthermore, the platform has not yet been tested on uneven terrain, leaving mobility challenges unresolved for future research iterations.
8Open Source Research Impact
Published on arXiv, this open source project invites makers to reproduce, modify, and investigate smartphone-based robotics. By providing a framework that reuses consumer electronics, the research offers a scalable path for academic and hobbyist robotics. Future developments may address arm integration and terrain adaptability to expand the platform's functional utility.