Hochschule Esslingen, Germany · 2026
Quadruped Locomotion
Reinforcement Learning Policy Integration on the Unitree Go2
Connecting policies to robot software
During an independent study / research period at Hochschule Esslingen, I worked with locomotion policy pipelines for the Unitree Go2. The work focused on the software interfaces between observations, policy inference, and robot commands.
The stack included ROS 2 Humble, PyTorch, Isaac Lab / Isaac Sim, and Unitree’s robot software interface.
System architecture & policy inference
I created and integrated ROS 2 observation and command pipelines, including messages for policy observations, desired low-level joint targets, and status. Low-level Unitree command bridging was an area of planned and ongoing integration work.
The final report will provide verified topic names, message definitions, and policy configuration.
Simulation-to-robot integration
The objective was to connect a learned locomotion policy to a robot-facing software pipeline. Command bridging, timing, and interface consistency are central integration concerns.
Hardware command limits, startup procedures, and safety checks must be verified against the report before they are described as implemented safeguards.
Evaluation & completion boundaries
The software pipeline work is described above. The final report and supporting media have not yet been supplied to verify end-to-end results. Successful closed-loop physical Go2 locomotion is not claimed.
Primary research document
Full research report
The compiled LaTeX report will be readable here and directly accessible as a PDF.
The compiled research PDF will be available here, with direct open and download links.
Demonstrations & resources
Approved robot photos, videos, and public code links will be added when available.