LocoMuJoCo

https://loco-mujoco.readthedocs.io/en/latest/
Admin
LocoMuJoCo is an imitation learning benchmark tailored for whole-body control, featuring diverse...
Tokens:198,279
Snippets:1,056
Trust Score:9.6
Update:1 month ago
Show doc for...
Tokens:
Raw
### Example of setting all caches with home directory path

Source: https://loco-mujoco.readthedocs.io/en/latest/_sources/source/loco_mujoco.installation.rst.txt

Example command that sets the cache path to the user's home directory.

```bash
loco-mujoco-set-all-caches --path "$HOME/.loco-mujoco-caches"
```

--------------------------------

### _setup_goal(_spec, _goal_type, _goal_params)

Source: https://loco-mujoco.readthedocs.io/en/latest/source/core/mujoco_base.html

Setup the goal.

```APIDOC
## _setup_goal(_spec: MjSpec, _goal_type: str, _goal_params: Dict) → Tuple[MjSpec, Goal]

### Description
Setup the goal.

### Parameters
- **spec** (_MjSpec_) - Specification of the environment.
- **goal_type** (_str_) - Type of the goal.
- **goal_params** (_Dict_) - Parameters of the goal.

### Returns
- **Tuple[MjSpec, Goal]** - Modified specification and the Goal.

```

--------------------------------

### Install Dependencies for SMPL-H Conversion

Source: https://loco-mujoco.readthedocs.io/en/latest/_sources/source/loco_mujoco.amass_installation.rst.txt

Installs required dependencies for the manual SMPL-H model conversion method.

```bash
pip install 'numpy<1.23.0' chumpy tqdm pyyaml
```

--------------------------------

### Install CPU PyTorch

Source: https://loco-mujoco.readthedocs.io/en/latest/_sources/source/loco_mujoco.amass_installation.rst.txt

Installs the CPU version of PyTorch to avoid conflicts with JAX. Recommended before installing SMPL-related packages.

```bash
pip install torch --extra-index-url https://download.pytorch.org/whl/cpu
```

--------------------------------

### Install SMPL Packages

Source: https://loco-mujoco.readthedocs.io/en/latest/_sources/source/loco_mujoco.amass_installation.rst.txt

Installs the SMPL-related packages. Note: this may install the GPU version of PyTorch unless the CPU version was already installed.

```bash
pip install -e '.[smpl]'
```

--------------------------------

### Install JAX with CUDA 12 support

Source: https://loco-mujoco.readthedocs.io/en/latest/_sources/source/loco_mujoco.installation.rst.txt

Installs JAX with GPU support for CUDA 12. Use this to enable GPU acceleration.

```bash
pip install jax["cuda12"]
```

--------------------------------

### Clone and install LocoMuJoCo in editable mode

Source: https://loco-mujoco.readthedocs.io/en/latest/_sources/source/loco_mujoco.installation.rst.txt

Clones the repository and installs it in editable mode. Run from the loco-mujoco directory.

```bash
cd loco-mujoco
pip install -e .
```

--------------------------------

### Test SMPL Model Setup

Source: https://loco-mujoco.readthedocs.io/en/latest/_sources/source/loco_mujoco.amass_installation.rst.txt

Tests the SMPL model setup. Assumes the DanceDB dataset has been downloaded; other datasets can be tested by modifying the script.

```bash
python examples/replay_datasets/smpl_example.py
```

--------------------------------

### Set Storage Paths Using Home Directory

Source: https://loco-mujoco.readthedocs.io/en/latest/_sources/source/loco_mujoco.amass_installation.rst.txt

Example using the home directory to set the storage paths for AMASS and SMPL models.

```bash
mkdir $HOME/amass
loco-mujoco-set-amass-path --path $HOME/amass
mkdir $HOME/smpl
loco-mujoco-set-smpl-model-path --path $HOME/smpl
```

--------------------------------

### Set Converted AMASS Path Example

Source: https://loco-mujoco.readthedocs.io/en/latest/_sources/source/loco_mujoco.amass_installation.rst.txt

Example using the home directory to set the converted AMASS data path.

```bash
loco-mujoco-set-conv-amass-path --path $HOME/amass_conv
```

--------------------------------

### Example usage of ImitationFactory with datasets

Source: https://loco-mujoco.readthedocs.io/en/latest/_sources/source/loco_mujoco.installation.rst.txt

Creates an imitation environment for UnitreeH1 using default and LAFAN1 datasets, then plays a trajectory. AMASS dataset configuration is commented out.

```python
from loco_mujoco.task_factories import ImitationFactory, LAFAN1DatasetConf, DefaultDatasetConf, AMASSDatasetConf

env = ImitationFactory.make(
    "UnitreeH1",
    default_dataset_conf=DefaultDatasetConf(["squat"]),
    lafan1_dataset_conf=LAFAN1DatasetConf(["dance2_subject4", "walk1_subject1"]),
    # amass_dataset_conf=AMASSDatasetConf([...])
)

env.play_trajectory(n_episodes=3, n_steps_per_episode=500, render=True)
```

--------------------------------

### create

Source: https://loco-mujoco.readthedocs.io/en/latest/source/loco_mujoco.task_factories.html

Creates an imitation learning environment with the specified configuration and preloads the trajectory.

```APIDOC
## create

### Description
Creates an imitation learning environment with the specified configuration and preloads the trajectory.

### Method
create

### Parameters
- **env_name** (_str_) - Required - The name of the registered environment to create.
- **default_dataset_conf** (_DefaultDatasetConf_, optional) - The configuration for the default trajectory.
- **amass_dataset_conf** (_AMASSDatasetConf_, optional) - The configuration for the AMASS trajectory.
- **lafan1_dataset_conf** (_LAFAN1DatasetConf_, optional) - The configuration for the LAFAN1 trajectory.
- **custom_dataset_conf** (_CustomDatasetConf_, optional) - The configuration for a custom trajectory.
- **terminal_state_type** (_str_, optional) - The terminal state handler to use. Defaults to "RootPoseTrajTerminalStateHandler".
- **init_state_type** (_str_, optional) - The initial state handler to use. Defaults to "TrajInitialStateHandler".
- **kwargs** - Additional keyword arguments to pass to the environment constructor.

### Returns
An instance of the requested imitation learning environment with the trajectory preloaded.

### Return Type
LocoEnv

### Raises
- **ValueError** - If the dataset_type is unknown.

```

--------------------------------

### Group observations with ObservationType and get group masks

Source: https://loco-mujoco.readthedocs.io/en/latest/source/tutorials/customizing_environments.html

Assign observations to groups (e.g., "prioritized" and "policy") in the observation_spec, then retrieve group masks via env.obs_container.get_obs_ind_by_group() and index the reset observation vector. The example creates a ToddlerBot environment with a walk task and calls env.create_observation_summary() to open a browser summary. Use this when your algorithm needs different observation subsets for different parts (e.g., policy vs. critic).

```python
import numpy as np
import jax

from loco_mujoco.core import ObservationType
from loco_mujoco import ImitationFactory


observation_spec = [
    ObservationType.FreeJointPosNoXY(obs_name="free_joint", xml_name="root", group="prioritized"),
    ObservationType.FreeJointVel(obs_name="free_joint_vel", xml_name="root", group="prioritized"),
    ObservationType.JointPos(obs_name="joint_pos", xml_name="left_hip_pitch", group=["prioritized", "policy"]),
    ObservationType.JointVel(obs_name="joint_vel1", xml_name="right_hip_pitch", group=["prioritized", "policy"]),
    ObservationType.JointVel(obs_name="joint_vel2", xml_name="left_knee", group=["prioritized", "policy"]),
    ObservationType.BodyPos(obs_name="head_pos", xml_name="head", group=["prioritized", "policy"]),
    ObservationType.LastAction(obs_name="last_action", group=["prioritized", "policy"]),
    # define many more in the order you want ...
]

env = ImitationFactory.make("ToddlerBot", observation_spec=observation_spec,
                            default_dataset_conf=dict(task="walk"))

# checkout the detailed observation space (opens webbrowser)
env.create_observation_summary()

# get observation masks
policy_mask = env.obs_container.get_obs_ind_by_group("policy")
prioritized_mask = env.obs_container.get_obs_ind_by_group("prioritized")

# get observations
obs = env.reset()
prioritized_obs = obs[prioritized_mask]
policy_obs = obs[policy_mask]
```

--------------------------------

### Initialize MyoSkeleton model

Source: https://loco-mujoco.readthedocs.io/en/latest/_sources/source/loco_mujoco.installation.rst.txt

Command to accept the license and download the MyoSkeleton model. Required to run the MyoSkeleton environment.

```bash
loco-mujoco-myomodel-init
```

--------------------------------

### ImitationFactory.make(env_name: str, default_dataset_conf: DefaultDatasetConf | Dict | DictConfig = None, amass_dataset_conf: AMASSDatasetConf | Dict | DictConfig = None, lafan1_dataset_conf: LAFAN1DatasetConf | Dict | DictConfig = None, custom_dataset_conf: CustomDatasetConf | Dict | DictConfig = None, terminal_state_type: str = 'RootPoseTrajTerminalStateHandler', init_state_type: str = 'TrajInitialStateHandler', **kwargs) -> LocoEnv

Source: https://loco-mujoco.readthedocs.io/en/latest/source/loco_mujoco.task_factories.html

Creates and returns an imitation learning environment given different configurations.

```APIDOC
## ImitationFactory.make(env_name: str, default_dataset_conf: DefaultDatasetConf | Dict | DictConfig = None, amass_dataset_conf: AMASSDatasetConf | Dict | DictConfig = None, lafan1_dataset_conf: LAFAN1DatasetConf | Dict | DictConfig = None, custom_dataset_conf: CustomDatasetConf | Dict | DictConfig = None, terminal_state_type: str = 'RootPoseTrajTerminalStateHandler', init_state_type: str = 'TrajInitialStateHandler', **kwargs) -> LocoEnv

### Description
Creates and returns an imitation learning environment given different configurations.

### Method
classmethod

### Parameters
#### Parameters
- **env_name** (str) - Required - The name of the environment to create.
- **default_dataset_conf** (DefaultDatasetConf | Dict | DictConfig, optional) - Configuration for the default dataset.
- **amass_dataset_conf** (AMASSDatasetConf | Dict | DictConfig, optional) - Configuration for the AMASS dataset.
- **lafan1_dataset_conf** (LAFAN1DatasetConf | Dict | DictConfig, optional) - Configuration for the LAFAN1 dataset.
- **custom_dataset_conf** (CustomDatasetConf | Dict | DictConfig, optional) - Configuration for a custom dataset.
- **terminal_state_type** (str, optional) - The terminal state handler to use. Defaults to "RootPoseTrajTerminalStateHandler".
- **init_state_type** (str, optional) - The initial state handler to use. Defaults to "TrajInitialStateHandler".
- **kwargs** - Additional keyword arguments.

### Returns
- **LocoEnv** - An instance of the requested imitation learning environment.
```

--------------------------------

### BoosterT1 constructor with default and custom specifications

Source: https://loco-mujoco.readthedocs.io/en/latest/_modules/loco_mujoco/environments/humanoids/boostert1.html

Initializes the BoosterT1 environment. If no spec is provided, uses the default XML file. Loads the MjSpec, sets default observation and action specifications if not provided, defaults to PD control with specified gains, and optionally modifies the spec for MJX if enabled.

```python
    def __init__(self, spec: Union[str, MjSpec] = None,
                 observation_spec: List[Observation] = None,
                 actuation_spec: List[str] = None, **kwargs):
        """
        Constructor.

        Args:
            spec (Union[str, MjSpec]): Specification of the environment.
                It can be a path to the xml file or a MjSpec object. If none, is provided, the default xml file is used.
            observation_spec (List[Observation]): Observation specification.
            actuation_spec (List[str]): Action specification.
            **kwargs: Additional arguments

        """

        if spec is None:
            spec = self.get_default_xml_file_path()

        # load the model specification
        spec = mujoco.MjSpec.from_file(spec) if not isinstance(spec, MjSpec) else spec

        # get the observation and action specification
        if observation_spec is None:
            # get default
            observation_spec = self._get_observation_specification(spec)
        else:
            # parse
            observation_spec = self.parse_observation_spec(observation_spec)
        if actuation_spec is None:
            actuation_spec = self._get_action_specification(spec)

        # uses PD control by default
        if "control_type" not in kwargs.keys():
            kwargs["control_type"] = "PDControl"
            kwargs["control_params"] = dict(p_gain=self.p_gains, d_gain=self.d_gains, scale_action_to_jnt_limits=False)

        # modify the specification if needed
        if self.mjx_enabled:
            spec = self._modify_spec_for_mjx(spec)

        super().__init__(spec=spec, actuation_spec=actuation_spec, observation_spec=observation_spec, **kwargs)
```

--------------------------------

### RLFactory.make(env_name: str, init_state_type: str = 'DefaultInitialStateHandler', terminal_state_type: str = 'HeightBasedTerminalStateHandler', goal_type: str = 'GoalRandomRootVelocity', reward_type: str = 'TargetVelocityGoalReward', **kwargs) -> LocoEnv

Source: https://loco-mujoco.readthedocs.io/en/latest/source/loco_mujoco.task_factories.html

Creates and returns an RL environment based on the specified environment name. Raises KeyError if the specified environment is not registered in Mujoco.registered_envs.

```APIDOC
## RLFactory.make(env_name: str, init_state_type: str = 'DefaultInitialStateHandler', terminal_state_type: str = 'HeightBasedTerminalStateHandler', goal_type: str = 'GoalRandomRootVelocity', reward_type: str = 'TargetVelocityGoalReward', **kwargs) -> LocoEnv

### Description
Creates and returns an RL environment based on the specified environment name.

### Method
static

### Parameters
#### Parameters
- **env_name** (str) - Required - The name of the registered environment to create.
- **init_state_type** (str, optional) - The initial state handler to use. Defaults to "DefaultInitialStateHandler".
- **terminal_state_type** (str, optional) - The terminal state handler to use. Defaults to "HeightBasedTerminalStateHandler".
- **goal_type** (str, optional) - The goal handler to use. Defaults to "GoalRandomRootVelocity".
- **reward_type** (str, optional) - The reward handler to use. Defaults to "TargetVelocityGoalReward".
- **kwargs** - Additional keyword arguments to pass to the environment constructor.

### Returns
- **LocoEnv** - An instance of the requested RL environment.

### Raises
- **KeyError** - If the specified environment is not registered in Mujoco.registered_envs.
```

--------------------------------

### get_height_matrix

Source: https://loco-mujoco.readthedocs.io/en/latest/source/core/terrain.html

Gets the height matrix of the terrain.

```APIDOC
## get_height_matrix(_matrix_config: Dict[str, Any], _env: Any, _model: MjModel | Model, _data: MjData | Data, _carry: Any, _backend: ModuleType) -> ndarray | Array

### Description
Get the height matrix of the terrain. This method returns the height matrix based on the provided configuration.

### Parameters
- **matrix_config** (_Dict[str, Any]_) - The matrix configuration.
- **env** (_Any_) - The environment instance.
- **model** (_Union_[MjModel, Model]) - The simulation model.
- **data** (_Union_[MjData, Data]) - The simulation data.
- **carry** (_Any_) - Carry instance with additional state information.
- **backend** (_ModuleType_) - Backend module used for computation (e.g., numpy or jax.numpy).

### Returns
- **Union[np.ndarray, jnp.ndarray]** - The height matrix of the terrain.

### Raises
- **NotImplementedError** - If the method is not implemented.
```

--------------------------------

### get_attribute_names

Source: https://loco-mujoco.readthedocs.io/en/latest/source/loco_mujoco.trajectory.html

Class method to get attribute names of TrajectoryModel.

```APIDOC
## get_attribute_names()

### Description
Returns the attribute names of the TrajectoryModel class.
```

--------------------------------

### Mjx class initialization

Source: https://loco-mujoco.readthedocs.io/en/latest/_modules/loco_mujoco/core/mujoco_mjx.html

Initializes the Mjx environment, optionally using MjWarp backend, and sets up the simulation model and data.

```python
class Mjx(Mujoco):
    """
    Base class for Mujoco environments using JAX.

    Args:
        n_envs (int): Number of environments to run in parallel.
        use_mjwarp (bool): If True, MjWarp will be used as the simulation backend instead of Mjx.
        nconmax (int): Maximum number of contacts to allocate for warp
        njmax (int): Maximum number of constraints to allocate for warp
        **kwargs: Additional arguments to pass to the Mujoco base class.

    """

    def __init__(self,
                 use_mjwarp=False,
                 nconmax=None,
                 njmax=None,
                 **kwargs):

        # call base mujoco env
        super().__init__(**kwargs)

        # add information to mdp_info
        self._mdp_info.mjx_env = True

        # time put model and data
        self._use_mjwarp = use_mjwarp
        if self._use_mjwarp:
            import warnings
            import time
            # todo: fix once mjwarp is stable
            warnings.warn(
                "mjwarp is experimental and may not work with all environments. "
                "Resetting is error prone. Training might work, but will not be optimal.",
                UserWarning
            )
            time.sleep(3)
            self.sys = mjx.put_model(self._model, impl='warp')
            self._first_data = mjx.make_data(self._model, impl='warp', nconmax=nconmax, njmax=njmax)
        else:
            self.sys = mjx.put_model(self._model)
            self._first_data = mjx.put_data(self._model, self._data)
```

--------------------------------

### get

Source: https://loco-mujoco.readthedocs.io/en/latest/source/loco_mujoco.trajectory.html

Retrieves data for a given trajectory and sub-trajectory index.

```APIDOC
## get

### Description
Retrieve the corresponding data for a given trajectory index and sub-trajectory index.

### Method
get

### Endpoint
get(_traj_index, _sub_traj_index, _backend)

### Parameters
- **traj_index** - Required - Index of the trajectory.
- **sub_traj_index** - Required - Index of the sub-trajectory.
- **backend** (ModuleType) - Optional - Backend to use for the computation.

### Returns
A new instance of SingleData containing the indexed data.
```

--------------------------------

### Create Virtual Environment for SMPL-H Conversion

Source: https://loco-mujoco.readthedocs.io/en/latest/_sources/source/loco_mujoco.amass_installation.rst.txt

Creates and activates a Python 3.10 virtual environment for manual SMPL-H model conversion. Includes a commented Windows activation command.

```bash
python3.10 -m venv smplh_model_conversion_env
source smplh_model_conversion_env/bin/activate

# On Windows:
# smplh_model_conversion_env\Scripts\activate
```

--------------------------------

### property free_jnt_qvel_id

Source: https://loco-mujoco.readthedocs.io/en/latest/source/core/mujoco_base.html

Gets the qvel index of free joints.

```APIDOC
## property free_jnt_qvel_id

### Description
Get the qvel index of free joints

### Returns
- **np.ndarray** - The qvel index of free joints with shape (n_free_joints, 6)

```

--------------------------------

### property free_jnt_qpos_id

Source: https://loco-mujoco.readthedocs.io/en/latest/source/core/mujoco_base.html

Gets the qpos index of free joints.

```APIDOC
## property free_jnt_qpos_id

### Description
Get the qpos index of free joints

### Returns
- **np.ndarray** - The qpos index of free joints with shape (n_free_joints, 7)

```

--------------------------------

### TaskFactory.get_name() -> str

Source: https://loco-mujoco.readthedocs.io/en/latest/source/loco_mujoco.task_factories.html

Gets the name of the current TaskFactory class.

```APIDOC
## TaskFactory.get_name() -> str

### Description
Gets the name of the current TaskFactory class.

### Method
classmethod

### Returns
- **str** - The name of the TaskFactory class.
```

--------------------------------

### _setup_observations

Source: https://loco-mujoco.readthedocs.io/en/latest/source/core/mujoco_base.html

Sets up the observation space for the environment. It generates a dictionary containing all the observation types and their corresponding information, as well as two dataclasses containing the indices in the Mujoco datastructure for each observation type (data_indices) and the indices for each observation type in the observation array (obs_indices).

```APIDOC
## _setup_observations

### Description
Sets up the observation space for the environment. It generates a dictionary containing all the observation types and their corresponding information, as well as two dataclasses containing the indices in the Mujoco datastructure for each observation type (data_indices) and the indices for each observation type in the observation array (obs_indices).

### Method
_setup_observations

### Endpoint
_setup_observations(observation_spec: List[ObservationType], model: MjModel, data: MjData) -> Tuple[ObservationContainer, ObservationIndexContainer, ObservationIndexContainer]

### Parameters
- **observation_spec** (List[ObservationType]) - Required - A list of observation types.
- **model** (MjModel) - Required - Mujoco model.
- **data** (MjData) - Required - Mujoco data structure.

### Returns
A dictionary containing all the observation types and their corresponding information, as well as two dataclasses containing the indices in the Mujoco datastructure for each observation type (data_indices) and the indices for each observation type in the observation array (obs_indices).
```

--------------------------------

### DomainRandomizer.get_name()

Source: https://loco-mujoco.readthedocs.io/en/latest/source/core/domain_randomizer.html

Class method to get the name of the domain randomizer class.

```APIDOC
## DomainRandomizer.get_name()

### Description
Get the name of the domain randomizer class.

### Method
classmethod

### Signature
`get_name() -> str`

### Returns
- **str** - The name of the domain randomizer class.
```

--------------------------------

### InitialStateHandler.get_name()

Source: https://loco-mujoco.readthedocs.io/en/latest/source/core/initial_state_handler.html

Class method to get the name of the initial state handler class.

```APIDOC
## InitialStateHandler.get_name()

### Description
Returns the name of the initial state handler class.

### Method
classmethod

### Signature
`get_name() -> str`

### Returns
- **str** - The name of the initial state handler class.

```

--------------------------------

### Create MJX environment with ImitationFactory and JAX vmap

Source: https://loco-mujoco.readthedocs.io/en/latest/source/tutorials/creating_environments.html

Creates an MJX environment (MjxUnitreeG1) with the stepinplace1 task using ImitationFactory. Uses JAX vmap and jit for parallel reset, step, and action sampling across 100 environments. Renders in parallel and logs steps per second every 100000 steps. The XLA_FLAGS environment variable can increase speed by ~30% on some GPUs.

```python
import os
import jax
import time

from loco_mujoco import ImitationFactory


# can increase the speed by ~30% on some GPUs
os.environ['XLA_FLAGS'] = (
    '--xla_gpu_triton_gemm_any=True ')

# create env
env = ImitationFactory.make("MjxUnitreeG1", disable_arms=True, default_dataset_conf=dict(task="stepinplace1"), use_mjwarp=False)

# create keys
key = jax.random.key(0)
n_envs = 100
keys = jax.random.split(key, n_envs + 1)
key, env_keys = keys[0], keys[1:]

# jit and vmap all functions needed
rng_reset = jax.jit(jax.vmap(env.mjx_reset))
rng_step = jax.jit(jax.vmap(env.mjx_step))
rng_sample_uni_action = jax.jit(jax.vmap(env.sample_action_space))

# reset env
state = rng_reset(env_keys)

step = 0
previous_time = time.time()
LOGGING_FREQUENCY = 100000
i = 0
while i < 100000:

    # step
    keys = jax.random.split(key, n_envs + 1)
    key, action_keys = keys[0], keys[1:]
    action = rng_sample_uni_action(action_keys)
    state = rng_step(state, action)

    # parallel render
    env.mjx_render(state)

    step += n_envs

    # log speed (disable rendering for accurate speed measurement)
    if step % LOGGING_FREQUENCY == 0:
        current_time = time.time()
        print(f"{int(LOGGING_FREQUENCY / (current_time - previous_time))} steps per second.")
        previous_time = current_time

    i+=1
```

--------------------------------

### Apptronik Apollo environment initialization and default observation specification

Source: https://loco-mujoco.readthedocs.io/en/latest/_modules/loco_mujoco/environments/humanoids/apptronik_apollo.html

Initializes the Apptronik Apollo environment by loading the MuJoCo model, setting default PD control, and building the observation specification. The first block handles spec loading, observation/action defaults, and MJX modification; the second block defines the default joint-position observations for the humanoid.

```python
if spec is None:
    spec = self.get_default_xml_file_path()

# load the model specification
spec = mujoco.MjSpec.from_file(spec) if not isinstance(spec, MjSpec) else spec

# get the observation and action specification
if observation_spec is None:
    # get default
    observation_spec = self._get_observation_specification(spec)
else:
    # parse
    observation_spec = self.parse_observation_spec(observation_spec)
if actuation_spec is None:
    actuation_spec = self._get_action_specification(spec)

# uses PD control by default
if "control_type" not in kwargs.keys():
    kwargs["control_type"] = "PDControl"
    kwargs["control_params"] = dict(p_gain=self.p_gains, d_gain=self.d_gains, scale_action_to_jnt_limits=False)

# modify the specification if needed
if self.mjx_enabled:
    spec = self._modify_spec_for_mjx(spec)

super().__init__(spec=spec, actuation_spec=actuation_spec, observation_spec=observation_spec, **kwargs)
```

```python
observation_spec = [
    # ------------- JOINT POS -------------
    ObservationType.FreeJointPosNoXY("q_floating_base", xml_name="floating_base"),
    ObservationType.JointPos("q_neck_yaw", xml_name="neck_yaw"),
    ObservationType.JointPos("q_neck_roll", xml_name="neck_roll"),
    ObservationType.JointPos("q_neck_pitch", xml_name="neck_pitch"),
    ObservationType.JointPos("q_torso_pitch", xml_name="torso_pitch"),
    ObservationType.JointPos("q_torso_roll", xml_name="torso_roll"),
    ObservationType.JointPos("q_torso_yaw", xml_name="torso_yaw"),
    ObservationType.JointPos("q_l_hip_ie", xml_name="l_hip_ie"),
    ObservationType.JointPos("q_l_hip_aa", xml_name="l_hip_aa"),
    ObservationType.JointPos("q_l_hip_fe", xml_name="l_hip_fe"),
    ObservationType.JointPos("q_l_knee_fe", xml_name="l_knee_fe"),
    ObservationType.JointPos("q_l_ankle_ie", xml_name="l_ankle_ie"),
    ObservationType.JointPos("q_l_ankle_pd", xml_name="l_ankle_pd"),
    ObservationType.JointPos("q_r_hip_ie", xml_name="r_hip_ie"),
    ObservationType.JointPos("q_r_hip_aa", xml_name="r_hip_aa"),
    ObservationType.JointPos("q_r_hip_fe", xml_name="r_hip_fe"),
    ObservationType.JointPos("q_r_knee_fe", xml_name="r_knee_fe"),
    ObservationType.JointPos("q_r_ankle_ie", xml_name="r_ankle_ie"),
    ObservationType.JointPos("q_r_ankle_pd", xml_name="r_ankle_pd"),
    ObservationType.JointPos("q_l_shoulder_aa", xml_name="l_shoulder_aa"),
    ObservationType.JointPos("q_l_shoulder_ie", xml_name="l_shoulder_ie"),
    ObservationType.JointPos("q_l_shoulder_fe", xml_name="l_shoulder_fe"),
    ObservationType.JointPos("q_l_elbow_fe", xml_name="l_elbow_fe"),
    ObservationType.JointPos("q_l_wrist_roll", xml_name="l_wrist_roll"),
    ObservationType.JointPos("q_l_wrist_yaw", xml_name="l_wrist_yaw"),
    ObservationType.JointPos("q_l_wrist_pitch", xml_name="l_wrist_pitch"),
    ObservationType.JointPos("q_r_shoulder_aa", xml_name="r_shoulder_aa"),
    ObservationType.JointPos("q_r_shoulder_ie", xml_name="r_shoulder_ie"),
    ObservationType.JointPos("q_r_shoulder_fe", xml_name="r_shoulder_fe"),
    ObservationType.JointPos("q_r_elbow_fe", xml_name="r_elbow_fe"),
    ObservationType.JointPos("q_r_wrist_roll", xml_name="r_wrist_roll"),
    ObservationType.JointPos("q_r_wrist_yaw", xml_name="r_wrist_yaw"),
    ObservationType.JointPos("q_r_wrist_pitch", xml_name="r_wrist_pitch"),
]
```
Tokens:
Raw