Learning to Open and Traverse Doors with a Legged Manipulator (2409.04882v1)

Abstract: Using doors is a longstanding challenge in robotics and is of significant practical interest in giving robots greater access to human-centric spaces. The task is challenging due to the need for online adaptation to varying door properties and precise control in manipulating the door panel and navigating through the confined doorway. To address this, we propose a learning-based controller for a legged manipulator to open and traverse through doors. The controller is trained using a teacher-student approach in simulation to learn robust task behaviors as well as estimate crucial door properties during the interaction. Unlike previous works, our approach is a single control policy that can handle both push and pull doors through learned behaviour which infers the opening direction during deployment without prior knowledge. The policy was deployed on the ANYmal legged robot with an arm and achieved a success rate of 95.0% in repeated trials conducted in an experimental setting. Additional experiments validate the policy's effectiveness and robustness to various doors and disturbances. A video overview of the method and experiments can be found at youtu.be/tQDZXN_k5NU.

• The paper introduces a teacher-student RL controller that enables autonomous door opening and traversal without pre-programmed routines.
• It employs domain randomization and simulation-to-real transfer, allowing the ANYmal robot to handle diverse door dynamics effectively.
• The approach achieved a 95% success rate, highlighting its robustness and potential for practical robotic applications in human-centric environments.

Learning to Open and Traverse Doors with a Legged Manipulator

The paper "Learning to Open and Traverse Doors with a Legged Manipulator" by Mike Zhang, Yuntao Ma, Takahiro Miki, and Marco Hutter presents a sophisticated solution to the challenging robotic task of autonomously navigating through doors. The approach leverages a learning-based controller, trained using a teacher-student framework, to enable the ANYmal robot to handle both push and pull doors without pre-programmed routines or user guidance on door characteristics.

Methodology

The authors propose a reinforcement learning (RL)-based controller coupled with domain randomization to enhance the robot's interaction capabilities with a range of door types. The policy is trained in a simulation environment and subsequently transferred to real-world applications using direct deployment without fine-tuning. The training method incorporates two key phases:

1. Teacher Policy Training: The teacher is trained with privileged access to door properties, enabling rapid and effective learning of the desired task behaviors.
2. Student Policy Training: The student mimics the teacher's actions based solely on proprioceptive and exteroceptive data available during real-world deployment.

The simulation environment models door characteristics, such as hinge and handle properties, with a focus on diverse and realistic variations. The RL framework utilizes Proximal Policy Optimization (PPO), and the teacher-student approach ensures the student policy can generalize across different scenarios without explicit prior information about the doors.

Numerical Results and Performance

The controller demonstrated a high success rate of 95.0% in experimental trials, emphasizing robustness and efficacy. Detailed experiments highlighted the policy's capability to autonomously infer door opening direction and handle varying door dynamics and dimensions effectively. The recurrent neural network (RNN)-based architecture of the student policy facilitated the estimation of door properties, which is instrumental for real-time adaptability.

Implications and Future Work

The implications of this work extend to enhancing autonomous robotic access to human-centric environments. The ability to traverse doors autonomously can substantially increase the operational scope of legged robots in diverse applications, from household assistance to industrial automation.

The theoretical contributions underline the importance of integrating teacher-student training paradigms and domain randomization in developing robust robotic policies. The demonstrated robustness to unmodeled disturbances suggests potential for broader applications where external perturbations are common.

Future developments could focus on incorporating onboard sensing for more autonomous operation, refining force sensing for better handle manipulation, and expanding capabilities to handle a wider array of door handles and locking mechanisms.

Conclusion

The paper presents a meticulously engineered approach that combines RL with a teacher-student training methodology to enable the ANYmal robot to open and traverse doors autonomously. The impressive success rate and robustness highlight the practical applicability and potential of this method in enhancing robotic capabilities in human-centric environments.