Robotic catching has traditionally focused on single-handed systems, which are limited in their ability to handle larger or more complex objects. In contrast, bimanual catching offers significant potential for improved dexterity and object handling but introduces new challenges in coordination and control. In this paper, we propose a novel framework for learning dexterous bimanual catching skills using Heterogeneous-Agent Reinforcement Learning (HARL). Our approach introduces an adversarial reward scheme, where a throw agent increases the difficulty of throws-adjusting speed-while a catch agent learns to coordinate both hands to catch objects under these evolving conditions. We evaluate the framework in simulated environments using 15 different objects, demonstrating robustness and versatility in handling diverse objects. Our method achieved approximately a$2 x$increase in catching reward compared to single-agent baselines across$\mathbf{1 5}$diverse objects.

Motion retargeting between heterogeneous polymorphs with different sizes and kinematic configurations requires a comprehensive knowledge of (inverse) kinematics. Moreover, it is non-trivial to provide a kinematic independent general solution. In this study, we developed a cyclic three-phase optimization method based on deep reinforcement learning for human-robot motion retargeting. The motion retargeting learning is performed using refined data in a latent space by the cyclic and filtering paths of our method. In addition, the human- in-the-loop based three-phase approach provides a framework for the improvement of the motion retargeting policy by both quantitative and qualitative manners. Using the proposed C- 3PO method, we were successfully able to learn the motion retargeting skill between the human skeleton and motion of the multiple robots such as NAO, Pepper, Baxter and C-3PO.

Motion imitation is a fundamental communication skill for a robot; especially, as a nonverbal interaction with a human. Owing to kinematic configuration differences between the human and the robot, it is challenging to determine the appropriate mapping between the two pose domains. Moreover, technical limitations while extracting 3D motion details, such as wrist joint movements from human motion videos, results in significant challenges in motion retargeting. Explicit mapping over different motion domains indicates a considerably inefficient solution. To solve these problems, we propose a three-phase reinforcement learning scheme to enable a NAO robot to learn motions from human pose skeletons extracted from video inputs. Our learning scheme consists of three phases: (i) phase one for learning preparation, (ii) phase two for a simulation-based reinforcement learning, and (iii) phase three for a human-in-the-loop-based reinforcement learning. In phase one, embeddings of the motions of a human skeleton and robot are learned by an autoencoder. In phase two, the NAO robot learns a rough imitation skill using reinforcement learning that translates the learned embeddings. In the last phase, the robot learns motion details that were not considered in the previous phases by interactively setting rewards based on direct teaching instead of the method used in the previous phase. Especially, it is to be noted that a relatively smaller number of interactive inputs are required for motion details in phase three when compared to the large volume of training sets required for overall imitation in phase two. The experimental results demonstrate that the proposed method improves the imitation skills efficiently for hand waving and saluting motions obtained from NTU-DB.

In this paper, an object detection and localization method for bin picking of plastic wrapped objects is described. Since such objects are deformable and have non-Lambertian surfaces, it is difficult to apply conventional feature point approaches or edge based template matching. To solve this problem, we propose a new method which is called “KCS (kernel convolution score)”. It measures the total score of convolution between local image and multiple kernels which are generated according to the characteristics of the target object. The local image having maximum score is considered as the most promising object for picking. In the last part of this paper, the limitation of existing edge and template matching in the given problem is discussed with the experimental results. Performance evaluation shows that the proposed method could localize the object which is appropriate for picking up with 91.67% success rate.