The paper "Learning Force Control for Legged Manipulation" by Tifanny Portela, Gabriel B. Margolis, Yandong Ji, and Pulkjit Agrawal presents a method for training reinforcement learning (RL) policies to control the force applied at the end effector of a legged manipulator. The authors aim to enable compliant interaction and whole-body force application, allowing teleoperators to modulate the gripper's compliance during locomotion and manipulation tasks. The method does not require access to force sensors but relies on proprioceptive sensing to estimate and modulate forces. The system is tested on a quadruped robot with an arm, demonstrating gravity compensation and impedance control. The learned whole-body controller allows humans to teleoperate the robot by commanding the manipulator, with the robot's body automatically adjusting to achieve the desired position and force. This enables a wide range of loco-manipulation tasks, such as pouring water and lifting objects, while maintaining compliance. The paper includes a detailed description of the system architecture, training process, and evaluation of force control performance in both simulation and real-world experiments. The results show that the learned policy can track forces with high accuracy, even in complex scenarios, and achieve significant workspace expansion through whole-body coordination. The authors also discuss the implications of their work for future research, including the potential for imitation learning and the exploration of more precise force control methods.The paper "Learning Force Control for Legged Manipulation" by Tifanny Portela, Gabriel B. Margolis, Yandong Ji, and Pulkjit Agrawal presents a method for training reinforcement learning (RL) policies to control the force applied at the end effector of a legged manipulator. The authors aim to enable compliant interaction and whole-body force application, allowing teleoperators to modulate the gripper's compliance during locomotion and manipulation tasks. The method does not require access to force sensors but relies on proprioceptive sensing to estimate and modulate forces. The system is tested on a quadruped robot with an arm, demonstrating gravity compensation and impedance control. The learned whole-body controller allows humans to teleoperate the robot by commanding the manipulator, with the robot's body automatically adjusting to achieve the desired position and force. This enables a wide range of loco-manipulation tasks, such as pouring water and lifting objects, while maintaining compliance. The paper includes a detailed description of the system architecture, training process, and evaluation of force control performance in both simulation and real-world experiments. The results show that the learned policy can track forces with high accuracy, even in complex scenarios, and achieve significant workspace expansion through whole-body coordination. The authors also discuss the implications of their work for future research, including the potential for imitation learning and the exploration of more precise force control methods.