LocoPilot is a teleoperation and data-collection framework for legged manipulators, built on a Unitree Go2 quadruped with a 6-DoF AgileX Piper L arm and two cameras. A second Piper L arm acts as the leader, and two handheld controllers let a single operator command the base, body posture, arm, and gripper concurrently. The system records synchronized multimodal observations and an 11D action vector that is shared between teleoperation and policy deployment.
Unlike robot-free interfaces, LocoPilot captures platform-specific commands, proprioception, and body-mounted observations directly on the target robot. The authors avoid a dedicated force/torque sensor by estimating follower arm interaction forces and rendering them to the operator through a passivity-constrained haptic channel. A ROS 2 pipeline handles data collection.
The authors evaluated LocoPilot across multiple loco-manipulation tasks and validated the collection-to-deployment pipeline by fine-tuning MolmoAct2 on pick-and-place demonstrations. The resulting policy ran on the physical robot without action retargeting or additional inverse kinematics, supporting the paper's claim that LocoPilot uniquely combines robot-in-the-loop legged data collection, sensor-free force feedback, and a shared action interface for deployment.