Learning Motor Skills of Reactive Reaching and Grasping of Objects

Reactive grasping of objects is an essential capability of autonomous robot manipulation, which is yet challenging to learn such sensorimotor control to coordinate coherent hand-finger motions and be robust against disturbances and failures. This work proposed a deep reinforcement learning based scheme to train feedback control policies which can coordinate reaching and grasping actions in presence of uncertainties. We formulated geometric metrics and task-orientated quantities to design the reward, which enabled efficient exploration of grasping policies. Further, to improve the success rate, we deployed key initial states of difficult hand-finger poses to train policies to overcome potential failures due to challenging configurations. The extensive simulation validations and benchmarks demonstrated that the learned policy was robust to grasp both static and moving objects. Moreover, the policy generated successful failure recoveries within a short time in difficult configurations and was robust with synthetic noises in the state feedback which were unseen during training.