This paper addresses the challenging task of twisting or removing lids from bottle-like objects using two multi-fingered robot hands. The authors train a control policy in simulation using deep reinforcement learning (RL) and then transfer it to a real-world setup with zero-shot learning. The policy is designed to handle complex bimanual manipulation tasks, which involve precise coordination and fine-grained contact between the robot's fingers and the object. Key contributions include: 1. **Physical Modeling**: A novel brake-based design is introduced to simulate the interaction between the lid and the bottle body, maintaining high fidelity while being computationally efficient. 2. **Perception**: A sparse object representation, extracted from off-the-shelf object segmentation and tracking tools, is sufficient for solving the perception problem, robust against occlusion and camera noise. 3. **Reward Design**: A keypoint-based contact reward is proposed to encourage natural lid-twisting behavior, which is crucial for learning the desired manipulation skills. The system is evaluated in both simulation and the real world, demonstrating successful lid twisting on a variety of objects with different physical properties. The learned policy generalizes well to unseen objects and is robust against perturbations, showcasing the effectiveness of the proposed approach in handling complex bimanual manipulation tasks.This paper addresses the challenging task of twisting or removing lids from bottle-like objects using two multi-fingered robot hands. The authors train a control policy in simulation using deep reinforcement learning (RL) and then transfer it to a real-world setup with zero-shot learning. The policy is designed to handle complex bimanual manipulation tasks, which involve precise coordination and fine-grained contact between the robot's fingers and the object. Key contributions include: 1. **Physical Modeling**: A novel brake-based design is introduced to simulate the interaction between the lid and the bottle body, maintaining high fidelity while being computationally efficient. 2. **Perception**: A sparse object representation, extracted from off-the-shelf object segmentation and tracking tools, is sufficient for solving the perception problem, robust against occlusion and camera noise. 3. **Reward Design**: A keypoint-based contact reward is proposed to encourage natural lid-twisting behavior, which is crucial for learning the desired manipulation skills. The system is evaluated in both simulation and the real world, demonstrating successful lid twisting on a variety of objects with different physical properties. The learned policy generalizes well to unseen objects and is robust against perturbations, showcasing the effectiveness of the proposed approach in handling complex bimanual manipulation tasks.