Teaching the parallel kinematic mechanisms to move using forward kinematic motion simulations and machine learning methods.
In this project, we developed and tested a machine learning-based control system for a parallel kinematic manipulator (pentapod). The short video below shows the experiment in action!
Using simulation data—generated by solving forward kinematics in an automated SolidWorks-based simulation—we recorded control parameters and end-effector poses. Multiple regression models were evaluated, and the best-performing one was selected for physical testing.
For real-world validation, we used ArUco marker tracking (OpenCV) to accurately measure the end-effector position. A synthetic test set (not used during training) ensured a fair evaluation.
Results:
✔️ Simulation accuracy: 93%
✔️ Physical prototype accuracy: 85%
(The difference stems from real-world complexities like friction, gravity, and mechanical tolerances.)
This work demonstrates that even for complex systems like parallel manipulators, machine learning can offer effective control solutions—paving the way for more precise and adaptive robotics in the future.