Grasping Unknown Objects with Only One Demonstration

The combination of imitation learning and reinforcement learning is expected to solve the challenge of grasping unknown objects with anthropomorphic hand-arm systems. However, this method requires a large number of perfect demonstrations and the implementation in real robots often differs greatly from the simulation effect. In this work, we introduce a curriculum learning mechanism and propose a multifinger grasping learning method that requires only one demonstration. First, a human remotely manipulates the robot via a wearable device to perform a successful grasping demonstration. The state of the object and the robot is recorded as the initial reference trajectory for reinforcement learning training. Then, by combining robot proprioception and the point cloud features of the target object, a multimodal deep reinforcement learning agent generates corrective actions for the reference demonstration in the synergy subspace of grasping and trains in simulation environments. Meanwhile, considering the topological and geometric variations of different objects, we establish a learning curriculum for objects to gradually improve the generalization ability of the agent, starting from similar to unknown objects. Finally, only successfully trained models are deployed on real robots. Compared to the baseline method, our method reduces dependence on the grasping data set while improving learning efficiency. Our success rate for grasping novel objects is higher.