Yongliang Wang, Hamidreza Kasaei
Interactive Robot Learning Laboratory (IRL-Lab)
2025 International Conference on Robotics and Automation (ICRA 2025)
Department of Artificial Intelligence, Bernoulli Institute, University of Groningen
Grasping large flat objects, such as books or keyboards lying horizontally, presents significant challenges for single-arm robotic systems, often requiring extra actions like pushing objects against walls or moving them to the edge of a surface to facilitate grasping. In contrast, dual-arm manipulation, inspired by human dexterity, offers a more refined solution by directly coordinating both arms to lift and grasp the object without the need for complex repositioning. In this paper, we propose a model-free deep reinforcement learning (DRL) framework to enable dual-arm coordination for grasping large flat objects. We utilize a large-scale grasp pose detection model as a backbone to extract high-dimensional features from input images, which are then used as the state representation in a reinforcement learning (RL) model. A CNN-based Proximal Policy Optimization (PPO) algorithm with shared Actor-Critic layers is employed to learn coordinated dual-arm grasp actions. The system is trained and tested in Isaac Gym and deployed to real robots. Experimental results demonstrate that our policy can effectively grasp large flat objects without requiring additional maneuvers. Furthermore, the policy exhibits strong generalization capabilities, successfully handling unseen objects. Importantly, it can be directly transferred to real robots without fine-tuning, consistently outperforming baseline methods.
We model the dual-arm task of grasping large flat objects as an MDP within a hierarchical RL framework, enabling the agent to learn actions directly from visual inputs. The architecture, shown in Figure 2, features an overhead camera mounted on the dual-arm robot that captures RGB-D images of the workspace. Each state consists of a color heightmap and a depth heightmap, generated by projecting the original RGB-D images along the gravity direction, and a mask heightmap that identifies the target object within the workspace. The task is divided into two sub-tasks: first, a large-scale grasp pose detection model processes input RGB images; second, a custom CNN-based RL model, trained with PPO, outputs designed primitive motions. This approach uses visual cues for feature extraction and depth data for action validation, improving decision-making. Unlike existing methods, our framework focuses on learning cooperative strategies between the two arms, with human-like actions.
We trained the proposed system using the Isaac Gym physics simulator and evaluated its performance in both simulated and real-world environments on a dual-arm robotic platform under various cluttered conditions. The simulated and real-world task setups are illustrated in Figure.Â
