← current work

Current work · Open source

Quad-SDK

I'm the maintainer of Quad-SDK, a full-stack, open-source, ROS-based framework for agile quadrupedal locomotion developed in CMU's Robomechanics Lab.

Quad-SDK vertically integrates planning, control, estimation, communication, and development tooling so researchers can go from simulation to hardware across many platforms with minimal changes. Its modular architecture lets you swap in your own implementation of any component while relying on the rest of the stack — and as maintainer I steward its architecture, releases, and the growing community around it (900+ GitHub stars).

Quad-SDK in action.

What's inside

  • Planning — a global body planner (RRT-Connect with mixed motion primitives), a nonlinear-MPC local planner, a Raibert-style footstep planner, and CBS-based multi-robot coordination.
  • Control & estimation — a real-time control loop, an Extended Kalman Filter for onboard state estimation, and a momentum-based external wrench observer, all behind a hardware-abstraction layer.
  • Learned controllers — support for training and deploying learned (reinforcement-learning) policies that drop into the same interface as the model-based controllers, in simulation and on hardware.
  • Simulation — Gazebo and MuJoCo backends (plus an NVIDIA IsaacSim beta), force-injection plugins, and multi-robot scenarios.
  • Tooling — logging & playback, batch and performance testing, and MATLAB/Python post-processing utilities.

Built on ROS with the Pinocchio dynamics library and released under the MIT license, Quad-SDK supports platforms including the Ghost Robotics Spirit and Vision60, Unitree Go1/Go2/A1/B2, and Boston Dynamics Spot.

Papers

Code Docs Video