HumanoidBench is a simulated benchmark for humanoid robots designed to evaluate and advance the field of humanoid robotics. It consists of 15 whole-body manipulation tasks and 12 locomotion tasks, such as shelf rearrangement, package unloading, and maze navigation. The benchmark aims to address the challenges of complex dynamics, sophisticated coordination among body parts, and long-horizon tasks. The benchmark is built on the MuJoCo physics engine and uses a Unitree H1 humanoid robot equipped with dexterous Shadow Hands. The benchmarking results show that state-of-the-art reinforcement learning (RL) algorithms struggle with most tasks, while a hierarchical learning approach, supported by robust low-level policies, achieves superior performance. HumanoidBench provides a platform for researchers to identify and address the challenges in humanoid robotics, facilitating the verification of algorithms and ideas. The open-source code is available at <https://humanoid-bench.github.io>.HumanoidBench is a simulated benchmark for humanoid robots designed to evaluate and advance the field of humanoid robotics. It consists of 15 whole-body manipulation tasks and 12 locomotion tasks, such as shelf rearrangement, package unloading, and maze navigation. The benchmark aims to address the challenges of complex dynamics, sophisticated coordination among body parts, and long-horizon tasks. The benchmark is built on the MuJoCo physics engine and uses a Unitree H1 humanoid robot equipped with dexterous Shadow Hands. The benchmarking results show that state-of-the-art reinforcement learning (RL) algorithms struggle with most tasks, while a hierarchical learning approach, supported by robust low-level policies, achieves superior performance. HumanoidBench provides a platform for researchers to identify and address the challenges in humanoid robotics, facilitating the verification of algorithms and ideas. The open-source code is available at <https://humanoid-bench.github.io>.