Implementation:Farama Foundation Gymnasium HumanoidEnv V5

Knowledge Sources	Farama_Foundation_Gymnasium Gymnasium Docs
Domains	Reinforcement_Learning, MuJoCo_Environments
Last Updated	2026-02-15 03:00 GMT

Overview

Concrete implementation of the Humanoid v5 MuJoCo locomotion environment provided by Gymnasium.

Description

The Humanoid environment is based on Tassa, Erez, and Todorov's work in "Synthesis and stabilization of complex behaviors through online trajectory optimization". The 3D bipedal robot simulates a human with a torso (abdomen), a pair of legs and arms, and tendons connecting hips to knees. The legs each consist of three body parts (thigh, shin, foot) and the arms consist of two body parts (upper arm, forearm). The goal is to walk forward as fast as possible without falling. The reward function combines a healthy reward, a forward reward based on center-of-mass x-velocity, a control cost penalty, and a contact cost penalty. The observation includes qpos, qvel, cinert, cvel, qfrc_actuator, and cfrc_ext data. The episode terminates if the humanoid's z-coordinate falls outside [1.0, 2.0].

Usage

Use this environment for benchmarking high-dimensional continuous control RL algorithms. It is one of the most challenging standard MuJoCo benchmarks due to its 17-dimensional action space and 348-dimensional observation space. The v5 version restores contact cost, excludes worldbody data from observations, and adds tendon info.

Code Reference

Source Location

Repository: Farama_Foundation_Gymnasium
File: gymnasium/envs/mujoco/humanoid_v5.py

Signature

class HumanoidEnv(MujocoEnv, utils.EzPickle):
    def __init__(
        self,
        xml_file: str = "humanoid.xml",
        frame_skip: int = 5,
        default_camera_config: dict[str, float | int] = DEFAULT_CAMERA_CONFIG,
        forward_reward_weight: float = 1.25,
        ctrl_cost_weight: float = 0.1,
        contact_cost_weight: float = 5e-7,
        contact_cost_range: tuple[float, float] = (-np.inf, 10.0),
        healthy_reward: float = 5.0,
        terminate_when_unhealthy: bool = True,
        healthy_z_range: tuple[float, float] = (1.0, 2.0),
        reset_noise_scale: float = 1e-2,
        exclude_current_positions_from_observation: bool = True,
        include_cinert_in_observation: bool = True,
        include_cvel_in_observation: bool = True,
        include_qfrc_actuator_in_observation: bool = True,
        include_cfrc_ext_in_observation: bool = True,
        **kwargs,
    )

Import

import gymnasium as gym
env = gym.make("Humanoid-v5")

I/O Contract

Inputs

Name	Type	Required	Description
action	np.ndarray (17,)	Yes	Torques applied to the 17 hinge joints (abdomen, hips, knees, shoulders, elbows), range [-0.4, 0.4]

Outputs

Name	Type	Description
observation	np.ndarray (348,)	State vector: qpos (22), qvel (23), cinert (130), cvel (78), qfrc_actuator (17), cfrc_ext (78); x,y excluded by default
reward	float	healthy_reward + forward_reward - ctrl_cost - contact_cost
terminated	bool	True if humanoid z-coordinate outside [1.0, 2.0]
truncated	bool	Episode truncation (handled by TimeLimit wrapper)
info	dict	Contains x_position, y_position, tendon_length, tendon_velocity, distance_from_origin, x_velocity, y_velocity, reward_survive, reward_forward, reward_ctrl, reward_contact

Usage Examples

import gymnasium as gym

env = gym.make("Humanoid-v5")
observation, info = env.reset(seed=42)

for _ in range(1000):
    action = env.action_space.sample()
    observation, reward, terminated, truncated, info = env.step(action)
    if terminated or truncated:
        observation, info = env.reset()

env.close()

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment