以CUDA技術實現高速視覺系統及具精準動態控制與力感測系統之擊球機器人

Abstract

此份研究使用高速雙眼相機來辨識投球者所投出的球，在視覺系統中，透過相機校正以及立體視覺，可將球體位置從二維影像平面轉換到三維空間，並且利用切換影像處理區域以及顯示卡的平行化運算，確保影像處理時間維持在相機取樣週期之中，實現短時間內之即時運算。 本論文考慮接近真實行為之球體飛行模型，加入重力以及空氣阻力的影響來預測飛行軌跡，並推導碰撞模型以及整合球體擊出至指定落點之各種條件，利用一最佳化方法求解擊球參數，並從個人電腦透過RS232將擊球參數傳輸至運動控制卡。 在機械手臂方面，利用Lagrange與 Newton-Euler方法，推導出機械手臂動態模型並加以驗證，接續使用參數估測法求得未知參數以描述完整動態，並實現控制器於高效能數位訊號處理器來即時規劃運動軌跡並擊球，使得手臂可以追隨所設定之路徑，在擊球時間點到達擊球位置以及所需之擊球速度。 最後於拍面加裝力感測器，擷取真實碰撞時間，驗證此時刻之系統控制效能。

This thesis is dedicated to the design and implementation of a four-joint robot in order to bat a flying ball to a designated location. The ball-batting robot uses a binocular camera to identify the position of a ball in the 3D space. Image processing is executed at 200 fps using a high-speed vision system and graphics processing unit (GPU). A flight model which takes into account the influence of gravity and air resistance is constructed to predict the ball’s flying trajectory. The estimation method is integrated with the collision model such that the contact point and specific timing for accomplishing the ball-batting task can be obtained by solving a constrained optimization problem. A 4-joint robotic arm is used in this experiment. The end-effector applied to rotate racket is a step motor controlled in open-loop, and the complete dynamic model of previous 3 joints is derived by Lagrange and Newton-Euler method, which is cross-validated by MATLAB Symbolic toolbox. The motion controller based on the new model is implemented on a high-performance DSP processor, and the predicted hit-point and velocity of the racket are sent from a personal computer through RS232 communication. Then the robot is commanded to strike the ball and send it to the desired location. Finally, a set of force sensors are mounted on the racket helping analyze the real collision timing and the corresponding system error.