LW-BenchHub

Overview

LW-BenchHub is a large-scale simulation framework built on Isaac-Lab Arena for training robots to perform common daily life tasks, with a current focus on kitchen manipulation and loco-manipulation tasks. The framework is developed by the Lightwheel team and provides comprehensive support for both teleoperation data collection and reinforcement learning training. Four key pillars underlie LW-BenchHub:

• High-quality and highly generalizable kitchen asset environments: featuring rich scene layouts, object placements, lighting conditions, and material diversity designed for teleoperation, RL, and IL tasks. The system supports seamless configuration, one-click switching between multiple robots (Unitree G1, PandaOmron, DoublePanda, LeRobot SO100/101 Arm, ARX-X7s, Agilex-Piper), various input devices (keyboard, VR, leader-follower arm), and diverse manipulation tasks to facilitate robust, reproducible research.
• Complete end-to-end data-to-policy pipeline: supporting the full sim2real loop: from teleoperation data collection, deterministic trajectory replay, to IL/RL model training, rigorous model evaluation, and smooth deployment from simulation to real-world robots.
• Highly efficient and heterogeneous training framework: Provides distributed training capabilities supporting both multi-machine multi-node and single-machine multi-node architectures. Enables zero-copy training within single-machine multi-node configurations, while supporting comprehensive environment isolation—allowing the model side and simulation side to operate in independent environments.

This documentation guide contains information about installation, getting started, and additional use cases such as accessing datasets, policy learning, and API docs.