s24_robosub_vip

S24 - Software

• Platform for running controller, algorithms, and sensors

  • Previously used Intel NUC to run all ROS nodes, but too slow for CV algorithms (including YOLOv8 model).
  • Jetson Orin Nano migration - GPU for processing YOLOv8 workloads.
    • Required development version of Jetson OS and ARM-compatible software dependencies.

• OAK-D integration - stereo camera for depth perception as well as higher resolution RGB sensor

  • Integration with software stack - difficult to publish images wirelessly to operator remote controlling the sub.
    • Only transmitted RBG data using Best Effort to maintain high frame rate.

• Position Controller - operates as a layer on top of existing velocity controller

  • Error calculation and PID control:
    • Calculates error between current position and target position in the world coordinate frame.
    • Runs PID calculations to generate a command velocity.
    • Position determined by pulling odometry information from the localization node.
  • Coordinate frame conversion:
    • The position controller generates velocity commands in the world coordinate frame, which must be converted to body coordinate frame for the velocity controller to use (e.g., forward 1 m/s is always in the positive-x direction in the body frame).
    • The conversion uses quaternion math (via numpy library) to accurately translate velocities and orientations between the two frames.
    • The velocity controller uses roll-pitch-yaw (Euler angles) internally.

• Path Following

  • A path follower node accepts a PoseArray containing waypoints which is fed in order to the position controller node.
  • The path follower node monitors feedback from position controller (i.e., when the sub is within a certain threshold from the target) to determine when to transition to the next waypoint.
  • This node doesn’t rely on odometry since that is handled by the position controller.

• Future work

  • Obstacle Perception
    • Using the OAK-D’s pointcloud simplifies obstacle perception, but it generates too many points (hundreds of thousands per image), making processing and transmission impractical.
    • We intend to send images to the mini-PC in lower resolution, reducing the pointcloud size.
    • This approach will enable detection of both object depth and color efficiently.
  • Controllers
    • Testing the position controller in real-world conditions was hindered by unresolved issues with the IMU and localization during water tests.
    • Tuning of PID constants is necessary for the controllers.
    • Currently, the controllers do not account for physical turn radius limits, potentially causing the submarine to receive commands it cannot execute.
    • Solutions include generating paths, such as Dubins curves, that respect these physical constraints to ensure feasible navigation.