Upkies are open-source wheeled biped robots. They have wheels for balancing and legs to negotiate uneven terrains. Upkies are designed to be buildable with off-the-shelf tools and components, like mjbots actuators. You can develop in Python or C++, on Linux or macOS, then deploy your behaviors to the robot's Raspberry Pi. Here are some instances of Upkies in action.

This repository contains all the materials needed to build and control an Upkie. Questions are welcome in the discussions forum or in the chat room.

Building your own Upkie

Step by step instructions to build a new Upkie from scratch are available in the Wiki.

Getting started

Upkies come with a model predictive controller that can balance and roam around. You can try it out in simulation by:

./start_mpc_balancer.sh

Once the agent is running, you can direct your Upkie using a gamepad 🎮

Left joystick: go forward right backward
Right joystick: turn left or right
Directional pad: down to crouch, up to stand up
Right button: (B on an Xbox controller, red circle on a PS4 controller) emergency stop 🚨 all motors will turn off

Click on the robot in the simulator window to apply external forces and see how the robot reacts.

Creating your own behaviors

Software for Upkies comes is packaged in an upkie Python library. You can install it:

From conda-forge: conda install -c conda-forge upkie
From PyPI: pip install upkie

When running on the real robot, your code will command the robot's actuators via another process called the spine. There are also simulation spines for testing before deploying to a robot. Let's start a Bullet simulation spine:

./start_simulation.sh

Now that we have a spine is running, we can control the robot in Python. For example:

import gymnasium as gym
import numpy as np
import upkie.envs
 
upkie.envs.register()
 
with gym.make("Upkie-Spine-Pendulum", frequency=200.0) as env:
    observation, _ = env.reset()
    gain = np.array([10.0, 1.0, 0.0, 0.1])
    for step in range(1_000_000):
        action = gain.dot(observation).reshape((1,))
        observation, reward, terminated, truncated, _ = env.step(action)
        if terminated or truncated:
            observation, _ = env.reset()

Other Gymnasium environments provide various levels of absraction to control the robot. They are listed in the Gym environments page of the documentation.

Going further

Examples

There are smaller standalone examples in the examples directory. For instance:

Domain randomization: shows how to add domain-randomization wrappers to an Upkie environment.
Lying genuflection: genuflect while lying on a horizontal floor.
Model predictive control: a self-contained MPC balancer
PD balancer: balance by proportional-derivative feedback to wheel velocities.

Some examples have optional dependencies, like those for the Genesis and PyBullet simulators. You can activate a virtual environment and install them as optional dependencies, or use Pixi:

pixi run --environment genesis ./examples/genesis_balancing.py

Tasks

Upkies come with a set of default behaviors that you can executed as Pixi tasks. To get started, make sure you have installed pixi.

Name	Task
`rlb3-enjoy-genesis`	Evaluate the last policy trained in Genesis
`rlb3-enjoy-pybullet`	Evaluate the last policy trained in PyBullet
`rlb3-train-genesis`	Train a new policy by reinforcement learning in Genesis
`rlb3-train-pybullet`	Train a new policy by reinforcement learning in PyBullet
`try-genesis`	Run a balancing example in Genesis
`try-pybullet`	Run a balancing example in PyBullet
`upkie-mpc-balancer`	Run the MPC balancer

You can execute a task by pixi run <task-name>, for instance:

pixi run upkie-mpc-balancer

Tasks are available both on your machine and on your Upkie's Raspberry Pi (Pixi comes pre-installed on the SD card image). They are implemented by agents. You can make your own agents by forking this repository or using the new_agent template to get started.

Contributing

Contributions are welcome to both the hardware and software of Upkies! Check out the contribution guidelines.

Citation

If you built an Upkie or use parts of this project in your works, please cite the project and its contributors:

@software{upkie,
  title = {{Upkie open source wheeled biped robot}},
  author = {Caron, St\'{e}phane and Perrin-Gilbert, Nicolas and Ledoux, Viviane and G\"{o}kbakan, \"{U}mit Bora and Raverdy, Pierre-Guillaume and Raffin, Antonin and Tordjman--Levavasseur, Valentin},
  url = {https://github.com/upkie/upkie},
  license = {Apache-2.0},
  version = {9.0.0},
  year = {2025}
}

Don't forget to add yourself to the BibTeX above and to CITATION.cff if you contribute to the project.