AgileX's Piper arm workflow enables imitation learning data collection in Isaac Lab

AgileX Robotics has released a comprehensive workflow for its Piper robotic arm that integrates with NVIDIA's Isaac Lab simulation framework to streamline teleoperation and data collection for embodied AI. The pipeline, built on ROS 2 Humble and Ubuntu 22.04, starts with keyboard-based teleoperation inside Isaac Lab, allowing developers to control the Piper arm for tasks like cube stacking. Manipulation trajectories are recorded in real-time, then processed into datasets compatible with Robomimic, a popular library for imitation learning. The workflow includes automatic conversion of the Piper arm's URDF model to USD format, gripper and actuator parameter configuration, and a custom observation pipeline for gripper state reading.

The project provides a fully functional external Isaac Lab environment with a manager-based task configuration, reward setup, and replay support for collected trajectories. This allows developers to rapidly iterate from human demonstration to reinforcement learning or behavior cloning without manual data preprocessing. The workflow is designed for researchers and engineers building robot learning applications, reducing the friction of simulation-to-real transfer. By providing a turnkey solution for data collection in a high-fidelity simulator, AgileX enables faster prototyping of generalist manipulation policies. The repository includes all necessary scripts, configuration files, and integration with AgileX's official URDF repository, making it accessible for labs working on imitation learning and reinforcement learning for arm manipulation.