Fast kNN-based prediction for the trajectory of a thrown body

Robotic throwing and catching of objects is a promising way of material transportation. For successful catch of the flying object accurate prediction of its trajectory in the gripper workspace is required. While most state-of-the-art solutions use physical models to get a trajectory forecast we apply a predictor based on nearest neighbor regression, which does not require exact physical model of the motion. The challenge of such application consist in high volume of calculations that are needed to compare the current trajectory with examples from the database. This issue is critical as the prediction must be real-time. Two approaches for speeding up the procedure are discussed. One approach is based on fast allocation of the small subset from the entire dataset. The current trajectory is compared only with the trajectories from this subset. Another approach is based on the parallelization of computations using graphical processing units. Both approaches are evaluated and compared based on real trajectories. The parallelized version of the algorithm is implemented on the robotic catching system. It provide successful catch for up to 86 % of thrown objects.