PhoneBot: A Bipedal Robot Controlled by a Repurposed Smartphone

Ask AI · From wheeled to bipedal, what new breakthroughs have mobile master control robots achieved?

“ $400, save to build a bipedal robot that can walk, fall and get up by itself, and also engage in dialogue.

A Huawei Honor 9 from 2017, lying in a drawer for nine years, now has a new job: to be the brain of a bipedal robot.

This robot is called PhoneBot, a new work from the RoMeLa laboratory in the Department of Mechanical and Aerospace Engineering at the University of California, Los Angeles (UCLA). It is 48.3 centimeters tall, weighs 1.8 kilograms, has 13 joints, and costs only $400.

According to the introduction, this robot can walk, chase people, take photos, engage in voice dialogue, and can even get up by itself using a push-up posture when it falls.

The most counterintuitive point is that there is no wire connecting the phone and the robot's body.

01.

Its 'brain' is hung outside the body via Wi-Fi.

Most robots on the market have their computing boards firmly welded into the body. To upgrade computing power, it's basically like rebuilding the entire machine.

PhoneBot takes a completely different approach. The phone is simply placed on the torso, with the camera, IMU, perception, and decision-making all handled by it; the generated motion commands are sent via Wi-Fi to the onboard single-board computer, which then drives 13 servos. There is no electrical connection between the phone and the body.

This design brings two obvious advantages.

First, replacing the 'brain' is extremely easy. The Honor 9 from 2017, and even cheap smartphones costing hundreds of yuan can handle it, just swap and use as needed.

The second is the direct restructuring of cost structure. High-definition cameras, IMU, CPU/GPU, wireless modules, and batteries—these are the most expensive components in robot perception and computing—have already been fully integrated in smartphones. PhoneBot is essentially riding for free on this integration. The budget saved is prioritized for the joint servo motors responsible for stepping and supporting the body.

In other words, the massive existing smartphone market has suddenly turned into a backup computing power pool for bipedal robots. On the order of several billion idle smartphones worldwide, in theory, all have the potential to become control units for robots.

And this idea has not appeared for the first time this time around.

02.

A cross-year research relay: from understanding the road, to standing on one's feet

In 2020, Intel's Intelligent Systems Lab released OpenBot. It was also about turning a smartphone into a robot, but OpenBot was deployed on a wheeled platform. A roughly $50 3D-printed wheeled chassis, with a smartphone plugged in via USB, responsible for seeing the road, planning paths, and following pedestrians.

The wheeled chassis has natural stability, high tolerance for signal delay and attitude disturbance, and this solution successfully completed the navigation task. This time, the RoMeLa laboratory research team from UCLA focused on “understanding the road” based on Further verification shows that smartphones can also help robots 'stand on their feet'.

PhoneBot still follows the main control logic of a smartphone, but makes a bold change: placing the phone on the torso without a wired connection, with all control commands transmitted wirelessly via Wi-Fi.

The phone takes on multiple roles, running simultaneously motion control, visual perception, and voice interaction three task lines.

• Motion is the top priority. It reads data from its built-in gyroscope and accelerometer, and synchronously collects the angles of each servo motor, calculating in real time how the legs should step, commanding the robot to move forward, turn, or get up after falling over;

• Visual recognition and localization are also completed locally on the phone, with the rear-facing camera responsible for identifying people and tracking targets, and using pose data to determine its own position, without the need for additional positioning modules;

• The voice function makes practical trade-offs, with the phone handling audio input and speaker output, while the computationally intensive large model dialogue and speech transcription require an external laptop for processing, with results sent back to the phone for output.

To ensure that simulated gait can be implemented on real hardware, the team pre-writes physical constraints such as servo motor torque limits and gear clearance in the simulation environment, avoiding a disconnect between virtual training and physical hardware; at the same time, they use step mirroring augmentation training techniques to address the issue of one-legged power bias in reinforcement learning.

Tested load data also confirms that the scheme has very low hardware requirements.

IMU data collection consumes 8.5%–13.4% CPU, gait calculation 6.3%–10.3%, target detection 15.6%–17.9%, and even the highest load visual localization only 25%–34.3%. This is why a 2017 Honor 9 is sufficient to support the entire bipedal control system.

03.

$400, rewriting the cost logic of quadruped robots

In addition to being able to walk on two legs and self-reset after falling, this robot's cost is even more surprising. The entire hardware system costs only $400, making the difference clear when compared with other bipedal platforms in the industry.

The cheapest category is educational kits costing tens of dollars, such as the ESP32, which can move two steps, but essentially is a moving teaching aid, not a motion learning platform. Hugging Face's LeRobot Humanoid this year, controlled by a Raspberry Pi, has parts costing about $2,500. Small-scale research bipedal platforms with autonomous balance and self-reset capabilities on the market Prices are generally in the thousands of dollars; many open-source bipedal kits also cost around a thousand to two thousand dollars.

PhoneBot is priced at $400, perfectly filling the gap between simple teaching aids and professional research prototypes.

Many small-scale bipedal platforms require separate acquisition of main controllers, cameras, and pose sensors, among other perception and computing components, adding up to a significant cost. PhoneBot takes a different approach, directly using the phone's built-in camera, IMU, and computing power to handle perception and gait calculation, no longer purchasing this set of perception and computing hardware separately, concentrating limited budgets on the most critical leg execution mechanisms for bipedal movement.

The ability to bring costs down to this level is not just due to reusing old phones. In recent years, sim-to-real (simulation to reality) training frameworks have gradually become open-source and mature, making gait training no longer an exclusive capability of a few laboratories, allowing ordinary researchers to complete replication using public solutions.

This round of cost reduction first benefits the research and education sectors, significantly lowering the entry threshold for bipedal gait research.

04.

What use is a robot that can't do anything?

Although the threshold has been lowered, from a practical application perspective, PhoneBot has no arms and has only been validated in flat terrain scenarios, with complex tasks like voice interaction still requiring external laptops for assistance. It is not a humanoid robot capable of doing housework or being practically useful.

Its true positioning is as a low-cost experimental prototype for gait research. Previously, doing bipedal motion research required expensive prototypes, and a single fall could easily damage parts, making debugging costly. Now, with just a $400 chassis and a used old phone, the entire solution is open-source and public. Falling, adjusting parameters, retraining, and trial-and-error costs are all significantly reduced.

For university research groups, student projects, and independent developers, this is a fundamental change.

From an industry perspective, the value of such low-cost open platforms lies in providing a referenceable architectural approach—reusing mature consumer hardware for perception and computing power, enabling more control ideas to iterate on physical hardware.

This combination of “consumption-end computing power + simple robot body”It also provides a low-cost route for lightweight embodied intelligence verification.

Project homepage:

https://arxiv.org/abs/2610.08737

https://arxiv.org/abs/2008.10631