wtf is sru-odin?

SRU x Odin takes a published robotics research method for guiding a walking robot around obstacles without a map, and makes it work on a real quadruped robot using a single sensor. The original research paper needed two separate pieces of hardware: a stereo camera to build a depth picture of the surroundings, and a separate tracking system to know precisely where the robot itself was. This project replaces both of those with one compact sensor called Odin1, which streams both the depth information and the robot's own position at once. The navigation approach works by feeding a depth image and a goal direction into a neural network that has a kind of short term memory, so it can recall where obstacles were a moment ago even as the robot keeps moving. That network was originally trained using reinforcement learning, a method where a simulated robot learns good behavior through repeated trial and error inside a physics simulator rather than being explicitly programmed with rules. Training happens inside a simulated environment, and the resulting trained model is then converted into a lightweight format that can run efficiently on the robot's onboard computer, without needing the full training software installed. The project targets the Unitree Go2 quadruped robot and outputs simple movement commands that any robot with the right control interface could use, so the navigation logic itself is not tied to one specific robot body. It also includes a written guide meant to be handed to an AI coding assistant, which can use that guide to regenerate the entire deployment software from scratch, covering the model interface, the robot control code, and the startup scripts. The repository is organized into three parts: one for training the navigation model inside a simulator, one for the tested deployment code that runs on the physical robot, and one containing the AI prompt guide. It is released under the MIT license and is based on a peer-reviewed research paper on mapless navigation.