Connecting to sources…

Upkie open-source wheeled biped · Self-balancing wheeled biped with wheels at the ends of its legs: wheels for balance, legs for uneven ground (6 mjbots brushless actuators, Raspberry Pi 4B)

Self-balancing wheeled biped with wheels at the ends of its legs: wheels for balance, legs for uneven ground (6 mjbots brushless actuators, Raspberry Pi 4B) · …

Maker
Open-source project started by Stéphane Caron (upkie)
Form
Self-balancing wheeled biped with wheels at the ends of its legs: wheels for balance, legs for uneven ground (6 mjbots brushless actuators, Raspberry Pi 4B)
For
Aimed, in the creator's words, at hobbyists and educators; the project's goals include education and training in model predictive control and reinforcement learning, and a research platform
Sensors
The IMU on the mjbots pi3hat drives attitude estimation and each actuator reports joint state; the base BOM has no camera or microphone, an add-on mounts an OAK‑D Lite depth camera, and the community has built an articulated head with a camera
Connectivity and app
Behaviours written in Python or C++ deploy straight to the on-board Raspberry Pi 4B; a USB gamepad (or a Bluetooth PS4 controller, per a guide) drives it; the pi3hat reaches the actuators over a CAN bus. Gymnasium environments run the same code in simulation and on the robot, and the docs mention no app or cloud service
Subscription
Open-source project, no subscription
Openness
The main upkie repo (software and build docs) is Apache‑2.0; the upkie_parts repo of 3D-printed parts (FreeCAD and STL) is Apache‑2.0; the upkie_description URDF repo is Apache‑2.0 except the wheel-tyre mesh, which is CC BY 4.0. The README and the creator's talk abstract say the mjbots actuators and boards it uses are open source in firmware, hardware (KiCad) and software
Privacy
The base build has no camera or microphone; control code runs locally on the on-board Raspberry Pi, and the docs mention no cloud service.
Not stated
Height and weight, battery life, itemised part prices, recommended age, step-by-step build docs for the v2 hardware

Vendor claims

  • Fully open hardware and software, buildable from off-the-shelf parts for around $3,000
  • Gymnasium environments let the same Python code control the simulator and the real robot
  • Made for hobbyists and educators, for teaching and training in model predictive control and reinforcement learning

Parts

  • 4 Model B · Runs robot software and control logic
  • pi3hat · Provides orientation sensing and CAN communication
  • mj5208 · Drives robot wheel rotation
  • qdd100 · Actuates hip and knee joints
  • qdd100 beta 3 · Drives leg joint motion
  • power dist, r4.3b+ · Distributes power to robot modules
  • PH-3 · Connects three-pin control and power wiring
  • pi3hat, r4.4b+ · Expands Raspberry Pi connections in the robot
  • XT-30 · Connects battery and main power wiring
  • XT90-S · Connects high-current power wiring
  • DS3231 · Provides a real-time clock and timekeeping
Parts and sources

Sources