I started this Project because i got inspired by some very smart People/Companies and Projects out there and want to understand and adapt their work. It is based on existing OpenSource-Projects and uses affordable Hardware to enable other people to build their own Bots and help us to understand how to control it the way we want.

This Project is heavily work in progress and may change every day. It is NOT a working or even finished Project. It is intended to be a community-project, so please feel invited to contribute.

See the first movements of SpotMicroAI on YouTube and the first contact with the NVIDIA Jetson Nano

Parts of this Project:

1. build a working physical Robot with cheap components everyone can build
2. create a simulated Environment and be able to control the Robot
3. to do RL training to make it learn how to stand/walk/run

First of all thanks to Kim Deok-Yeon aka KDY0523 who made this incredible work on Thingiverse

This basically is the physical Robot. It will take some days to print and assemble all the parts, but it's worth all the effort. I also sanded, primed and painted all the Parts to give it a nicer look.

Here is my Thingiverse-Make

Since my setup required some additional Hardware, i recreated some parts using FreeCAD - see /urdf/FreeCAD-Directory. You might notice the redunancy of some STL-files, which is caused by the "ROSification" and the slightly different Model-Structure. I will have to clean it up.

The Brain of all this is the NVIDIA JetsonNano. It has a 16 Channel PCA9685 I2C-Servo Driver connected, which is used to control the servos. The IMU (GY-521) is also connected via I2C and provides roll and pitch angles of the Robot. The OLED-Display is used to have some nice output. I will provide a Fritzing-Layout in the near future.

You can find the all the Code for the Jetson Nano here.

UPDATE: I discovered NVIDIA's Isaac SDK and will spend some time in exploring it.

SpotMicroAI has a RPi-Cam, Sonar-Sensors and an IMU Gyro/Accel-Sensor (as designed by KDY). In one of the first versions i added two additional Sonars at the bottom, but i had to remove them again to have space for the voltage regulators. TODO: We will need a Bottom-Case version 2 here.

The Rear-Part has space for an OLED-Display (SSD-1306) and a LED Power-Button. In a first version i used an Arduino Mega as kind of Servo/Sensor-Controller and a Raspberry PI as Locomotion-Controller (communication via UART). But it showed up that the Arduino is too slow to handle Sensor-Signals and Servo-PWM properly at the same time. Now i use the Jetson with the PCA as described above.

I am not sure if the Hardware i use now will be enough to finally have a very smooth walking Robot like for example the real SpotMini. See this more as a Research-Project where I try to use cheap Hardware and other People's Work to learn more about how this all works.

• I am still working on the PowerSupply. Currently SpotMicroAI is powered by two separate AC-Adapters for Jetson and PCA. I need another voltage-regulator for the Jetson (7,4v -> 5V).

Masses and Inertias of the URDF-Model are still not correct.

There is also a Blender-File included which i used to create the STLs for the simulation. Of course you could also do some nice renderings with it! :)

This example can be found in the Repository. You need a GamePad for this to work:

pip3 install numpy
pip3 install pybullet
pip3 install inputs
...TODO: provide setup.py

cd Core/
python3 example_automatic_gait.py

There is also a first ROSification of SpotMicroAI.

First of all install ROS. I use Melodic, but it should work with Kinetic, too. I will not go into detail on how to install ROS because there are many good Tutorials out there.

When finished installing ROS:

cd ~/catkin_ws/src
git clone https://github.com/FlorianWilk/SpotMicroAI.git
cd ..
catkin_make
source ./devel/setup.bash
roslaunch spotmicroai showmodel.launch

This will show up RVIZ with the Model of SpotMicroAI.

TODO: Next steps will be to create a ROS-Sim-Adapter to map the joint_states to the leg_topic and to create an Implementation for the Jetson which handles leg_topics -> Servos. Then i will write a Controller-Node, which uses the Kinematics/RL-Model to control the Bot via the leg_topics. This Controller-Node will just be a wrapper for the same logic we use for the PyBullet-Simulation.

I like the Ideas of

• this Paper by Jie Tan, Tingnan Zhang, Erwin Coumans, Atil Iscen, Yunfei Bai, Danijar Hafner, Steven Bohez, and Vincent Vanhoucke Google Brain,Google DeepMind
• this Paper by Michael Lutter, Christian Ritter & Jan Peters ∗
• and this one by Abhik Singla, Shounak Bhattacharya, Dhaivat Dholakiya, Shalabh Bhatnagar, Ashitava Ghosal, Bharadwaj Amrutur and Shishir Kolathaya
• also this one by Krzysztof Choromanski,Aldo Pacchiano,Jack Parker-Holder,Yunhao Tang,Deepali Jain,Yuxiang Yang,Atil Iscen,Jasmine Hsu,Vikas Sindhwani

In order to be able to move the Robot or even make it walk, we need something which tells us what servo-angles will be needed for a Leg to reach position XYZ. This is what InverseKinematics does. We know all the constraints, the length of the legs, how the joints rotate and where they are positioned.

You can find some a first draft of the calculations here. There is also a Jupyter Notebook explaining the Kinematics and a YouTube-Video.

There is no real Training-Code yet. I started to create a simple OpenAI-Gym-Environment, but have not finished it yet :/ TODO:

• Finish the OpenAI-Gym-Env
• write a basic RL-based Implementation to support example_automaticgait.py with a "BodyBalancer". ActionSpace is x,y,z of the Body, ObservationSpace pitch,roll,ground_distance,kinematic_motion_function_index
• adapt the "Neural Network Walker"-Example from Bullet to SpotMicroAI? not sure..

• Deok-yeon Kim creator of SpotMicro
• Boston Dynamics who built this incredible SpotMini,
• Ivan Krasin - https://ivankrasin.com/about/ - thanks for inspiration and chatting
• My colleagues at REWE digital / Research & Innovation for inspiration and feedback
• Jie Tan, Tingnan Zhang, Erwin Coumans, Atil Iscen, Yunfei Bai, Danijar Hafner, Steven Bohez, and Vincent Vanhoucke Google Brain,Google DeepMind