SpotMini Clone Will Use Machine and Reinforcement Learning for Kinematics and Motion

German engineer Florian Wilk is on a lofty endeavor to build a Boston Dynamics’ SpotMini clone that uses AI for the robots kinematics, leg and body motions, and he’s already well underway in achieving that goal. His SpotMicroAI robot is designed around several open source projects, including Deok-Yeon Kim’s Spotmicro, and employs affordable and readily available hardware. It’s important to note at this point that the robot is not currently functioning, and is a work in progress, in fact, Wilk is inviting anyone willing to work on this project to join him to help make it fully functional.

There are three parts to this project, with the first being to build the actual robot and get it fully functional. The second will be to design a simulated environment and be able to control the robot, and the third centers on implementing the machine/reinforcement learning, enabling it to stand, walk, maneuver, and navigate.

On the hardware end, the SpotMicoAI is outfitted with an NVIDIA Jetson Nano, plus a 16-channel PCA9685 I2C servo driver that will eventually be used to actuate the robot’s legs at some point. It’s also packing four HC-SR04 SonarSensors, with two in the front and two in the bottom, as well as an MPU-6050 MEMS MotionTracking IMU. There’s even a Raspberry Pi Camera that will be used for object detection, along with an RPLidar A1M8 3600 laser scanner. Florian states he may incorporate sonar sensors for object detection/avoidance instead of cameras in the near future.

For simulation and kinematics, Wilk is currently using PyBullet, but hasn’t adapted it for ML/RL training yet, and states he’s using Google’s paper- Sim-to-Real: Learning Agile Locomotion For Quadruped Robots, as a base to help him achieve his project goal. Those interested can keep track of the SpotMicroAI on Florian’s RobotShop and GitHub pages.