Iretiayo Akinola*, Jingxi Xu*, Shuran Song, and Peter Allen
Columbia Robotics Lab
Grasping in dynamic environments presents a unique set of challenges. A stable and feasible grasp can become infeasible as the target object moves; this problem becomes exacerbated when there are obstacles in the scene. One common approach is to switch between a set of pre-planned grasps as they become feasible but it usually results in large swinging arm trajectory motions as the closest grasp in the list may require a complete change in arm motion. This can result in failure especially if the object is moving relatively fast. In this work, our grasp planning approach is aware of the arm’s reachability and the object’s motion. We model the reachability space of the robot using a signed distance field and quickly screen the grasps in the database. Also, we train a neural network to predict the grasp quality conditioned on the current motion of the target. To move the arm to the planned reachable grasp, we use trajectory optimization with seeding for arm motion generation. This keeps the newly generated trajectory near-optimal and close to the previously generated arm motion to reduce the fluctuation. Our experiments on linear and circular motion demonstrates that our method outperforms the baseline methods significantly in picking fast moving targets within static obstacles.