中国科学院自动化研究所机构知识库

机器人抓取操作面临各种环境的干扰以及动态变化，尤其在密集杂乱环境中，物体位姿状态千变万化且相互遮挡，机器人仅依靠单一抓取操作往往难以完成抓取任务，多操作协同是提高抓取成功率的有效途径。本文针对密集杂乱场景下的抓取问题，开展基于视觉感知的多操作协同技术研究，从机器人吸-抓可供性学习、推-抓协同操作和面向杂质性目标的精准抓取方面开展讨论。本文的主要工作如下：
(1) 根据杂乱密集场景下操作动作效用互补作用，以抓取和吸取动作为动作原语，开展基于数据驱动的机器人操作协同抓取方法研究。通过针对每一种动作原语在典型场景下操作位置的精准标注，构建操作数据集；采用全卷积神经网络学习每个动作原语的像素级可供性图；分析不同动作原语的可供性图热度分布信息，设计启发式协同规划器，实现机器人吸-抓协同操作。实验结果展示了该方法在杂乱环境抓取任务中的有效性。
(2) 针对数据驱动方法训练成本高、泛化性较差的问题，提出了一种基于强化学习的推-抓协同操作方法。该方法将全卷积神经网络与Q-Learning方法相结合，通过环境交互实现操作点可供性Q值学习，结合操作协同机制，实现对最佳动作的选择。为提高推动动作效率，提出了一种基于物体分散度和知识引导的推动策略学习机制。该机制主动引导机器人分散物体，并结合先验知识，降低动作策略空间搜索范围，加快策略收敛速度。实验结果表明，所提推-抓协同方法降低了无效动作率，提高了策略训练效率，在随机场景和挑战性密集场景中均有较好的性能表现。
(3) 针对杂质性目标的抓取任务，提出了一种基于注意力机制的推-抓协同操作策略方法。该注意力机制包括了针对可见目标的显著性检测及定位，以及针对不可见目标的环境物体聚类分析及目标隐藏区域推理。为了提高面向目标的推动操作的有效性，提出了一种基于目标为中心分散度度量的主动推动策略，引导机器人将目标与其他物体分离。实验结果验证了该方法在杂乱密集环境中主动定位杂质性目标并完成抓取的高效性。

Robotic grasping is faced with various environmental interference and dynamic changes, especially in dense clutter where the poses of objects are changeable and mutually occluded. It is difficult for robot to complete the grasping task only relying on prehensile actions. Multi-manipulation synergy is an effective way to improve the success rate of grasping. This thesis focuses on the research of vision-based manipulation synergy technology in dense-clutter grasping task, and discusses the aspects of robot suction-grasp affordance learning, push-grasp synergy manipulation and precise grasping for impurity targets. The main contents of this thesis are as follows:

(1) According to mutual complement of different actions in dense clutter, a suction-grasp manipulation method based on data-driven modeling is proposed. The manipulation database is constructed by precisely annotating suction and grasp proposals over images from the real typical scenes. And the pixel-wise probability maps of the affordances for two primitive actions are learned through a fully convolutional neural network. With the analysis of affordance distribution, the synergy planner based on heuristic rules is designed to realize robot suction-grasp manipulation. The experimental results show the effectiveness of this method in dense-clutter grasping tasks. 

(2) A push-grasp manipulation method based on reinforcement learning is proposed to deal with the problem of high training cost and poor generalization of data-driven method. This method involves training fully convolutional networks that map from visual observations to actions in a Q-learning framework to learn the Q-value of manipulation points through environment interaction, and combines with manipulation synergy mechanism to realize the selection of the best action. To improve the efficiency of pushing action, a learning mechanism based on object dispersion degree metric and knowledge induction is proposed. The mechanism actively guides robot to disperse the objects and reduces the search range of the state space, which greately accelerate the policy convergence. The experimental results show that the proposed method lowers the invalid action rate, improves the efficiency of policy training, and has better performance in random scenes and challenging scenes. 

(3) Aiming at the grasping task of impurity targets, a push-grasp manipulation policy based on attention mechanism is proposed. The attention module includes the saliency detection and location for visible targets, and clustering analysis of environmental objects and reasoning of hidden regions for invisible objects. To improve the effectiveness of the target-oriented pushing action, an active pushing policy based on the target-centric dispersion degree is proposed to guide the robot to isolate the targets from other objects. The experimental results verify the effectiveness of our method in actively locating and grasping the impurity targets in dense clutter.