This paper presents a deep reinforcement learning (RL) approach for bimanual manipulation tasks, specifically for twisting lids off various bottle-like objects. The method involves training a policy in simulation using deep RL and then transferring it to the real world without any additional training. The policy is designed to handle a wide range of objects with different physical properties, including varying shapes, sizes, and materials. The system uses a novel physical model for articulated objects, which includes a brake link to simulate friction between the bottle body and lid. The policy is trained using a sparse object representation and a reward function that encourages natural contact and twisting behavior. The system also incorporates real-time perception using a depth camera and a segmentation model to track object parts. The policy is tested in both simulation and the real world, demonstrating its ability to generalize to a variety of objects and robustly handle perturbations. The results show that the policy can successfully twist lids off various objects, including household items, and outperforms baselines in terms of performance and generalization. The study highlights the effectiveness of sim-to-real transfer in complex bimanual manipulation tasks.This paper presents a deep reinforcement learning (RL) approach for bimanual manipulation tasks, specifically for twisting lids off various bottle-like objects. The method involves training a policy in simulation using deep RL and then transferring it to the real world without any additional training. The policy is designed to handle a wide range of objects with different physical properties, including varying shapes, sizes, and materials. The system uses a novel physical model for articulated objects, which includes a brake link to simulate friction between the bottle body and lid. The policy is trained using a sparse object representation and a reward function that encourages natural contact and twisting behavior. The system also incorporates real-time perception using a depth camera and a segmentation model to track object parts. The policy is tested in both simulation and the real world, demonstrating its ability to generalize to a variety of objects and robustly handle perturbations. The results show that the policy can successfully twist lids off various objects, including household items, and outperforms baselines in terms of performance and generalization. The study highlights the effectiveness of sim-to-real transfer in complex bimanual manipulation tasks.