How jlancha/convex-mpc-quad-sdk works
Project Type and Purpose
• This is a ROS-based, full-stack robotics framework (“Quad-SDK”) for agile quadrupedal locomotion in simulation and on hardware.
• It integrates planning (global & local), model-predictive control (linear MPC via QP), state estimation, communication, drivers, simulation (Gazebo), visualization (RViz), logging, and developer tools.
Top-Level Architecture Components
a. ROS Middleware (core “bus”)
• ROS Master / Parameter Server
• Topic, Service, and Action interfaces defined by quad_msgs
b. Perception & Estimation
• mocap_optitrack: Motion capture data → robot pose
• robot_driver/estimators: IMU, encoders → state estimator (EKF, complementary filter)
• quad_utils/terrain_map_publisher: Depth/mesh → terrain map
c. Planning
• global_body_planner: Fast global motion planner → /robot_x/global_plan
• local_planner: Footstep sequence planner → /robot_x/foot_plan
d. Control
• nmpc_controller / lmpc_controller: Generate optimized gait force/foot trajectories via OSQP-Eigen QP → leg commands
• robot_driver/controllers: PID, inverse dynamics → motor torque commands
e. Communication & Teleoperation
• quad_utils/remote_heartbeat: Heartbeat, remote supervision
• teleop_twist_keyboard / teleop_twist_joy: User cmd_vel → twist commands
f. Simulation & Hardware Interface
• quad_simulator: Gazebo URDF/SDF models + plugins (contact, estimator, controller)
• robot_driver/hardware_interfaces/spirit_interface: Real hardware abstraction
g. Utilities & Logging
• quad_logger: ROS bag reader, plot scripts
• quad_utils/rviz_interface, trajectory_publisher, mesh_to_grid_map, etc.
h. Examples & Third-party
• casadi_examples, communication_test, external/ (OSQP, RBDL-orb, grid_map_pcl, mavlink)
Component Interactions & Data Flows
• Sensor plugins (Gazebo) & mocap publish raw sensor topics → Estimator nodes → publish /robot_x/body_state
• Estimation and user teleop feed into Planners:
Global Planner consumes /body_state → publishes global path
Local Planner consumes global path + terrain map → publishes footstep plan
• Controller stack:
MPC (nmpc/lmpc) consumes /body_state, footstep plan → computes leg force/foot trajectories → publishes LegCommandArray
Motor controllers subscribe to LegCommand → compute low-level torques → publish to hardware interface or Gazebo plugin
• Utils nodes subscribe to topics for visualization (RViz), terrain mapping, logging.
• All inter-node comms via ROS topics/services.
Architectural Patterns & Principles
• Modular ROS packages for separation of concerns
• Layered pipeline: Sensing → Estimation → Planning → Control → Actuation
• Plug-and-play: Swap planners or controllers by remapping ROS topics
• External solver integration: OSQP via OSQP-Eigen for QP in MPC
• Simulation-hardware parity via common ROS interfaces