Environments

IGParkourEnv

The main Isaac Gym environment for motion tracking and parkour tasks.

Initialization

from parc.motion_tracker.envs.ig_parkour.ig_parkour_env import IGParkourEnv

env = IGParkourEnv(
    config=config,
    num_envs=4096,
    device="cuda:0",
    visualize=False
)

Parameters:

config (dict): Environment configuration containing:
- env: Environment-specific settings
- Character model, terrain, observation/action spaces
num_envs (int): Number of parallel environments
device (str): PyTorch device (“cuda:0” or “cpu”)
visualize (bool): Enable visualization and rendering

Source: ig_parkour_env.py:40-149

Key Attributes

env._num_envs              # Number of parallel environments
env._device                # PyTorch device
env._visualize             # Visualization enabled flag
env._num_dm_envs           # Number of DeepMimic environments
env._termination_height    # Height threshold for termination
env._episode_length        # Maximum episode length
env._timestep              # Simulation timestep duration

Observation Space

Observations include:

Root state: Position, rotation, velocity, angular velocity
Joint rotations: Character joint angles
DOF velocities: Joint velocities
Key body positions: End effector positions
Target observations: Future reference motion states (if enabled)
Contact information: Binary contact states (if enabled)
Heightmap: Local terrain height field

obs_space = env.get_obs_space()
obs = env._compute_obs()  # Returns [num_envs, obs_dim] tensor

Source: ig_parkour_env.py:842-965

Action Space

action_space = env.get_action_space()
# Returns gym.spaces.Box with shape [dof_size]

Actions are joint torques applied to character DOFs.

Core Methods

reset()

Resets specified environments to initial state.

obs, info = env.reset(env_ids=None)

Parameters:

env_ids (torch.Tensor, optional): Environment indices to reset. If None, resets all.

Returns:

obs (torch.Tensor): Observations [num_envs, obs_dim]
info (dict): Additional information

Source: ig_parkour_env.py:809-829

step()

Executes one simulation step with actions.

next_obs, reward, done, info = env.step(action)

Parameters:

action (torch.Tensor): Actions [num_envs, action_dim]

Returns:

next_obs (torch.Tensor): Next observations
reward (torch.Tensor): Rewards [num_envs]
done (torch.Tensor): Done flags [num_envs]
info (dict): Additional information including:
- rewards: Detailed reward components
- timestep: Current timestep per environment
- char_contact_forces: Contact forces on character

Source: ig_parkour_env.py:831-840

Reward Structure

Reward computation based on multiple components:

# Weighted reward components (ig_parkour_env.py:984-1044)
reward = (
    pose_w * pose_reward +          # Pose similarity
    vel_w * velocity_reward +       # Velocity matching
    root_pos_w * root_pos_reward +  # Root position tracking
    root_vel_w * root_vel_reward +  # Root velocity tracking
    key_pos_w * key_pos_reward      # End effector tracking
)

# Optional contact penalty
if use_contact_info:
    reward += contact_penalty

Reward Components:

pose_r: Joint rotation similarity to reference
vel_r: Joint velocity similarity
root_pos_r: Root position tracking error
root_vel_r: Root velocity tracking error
key_pos_r: Key body position tracking error
contact_penalty: Contact timing mismatch penalty

Source: ig_parkour_env.py:984-1044

Termination Conditions

Episode termination occurs when:

Height termination: Root height below threshold
Pose termination: Pose deviation exceeds limit (if enabled)
Root position termination: Root position error too large
Root rotation termination: Root rotation error too large
Time limit: Episode length exceeded
Motion completion: Reference motion finished

# Configuration (ig_parkour_env.py:62-88)
env_config = {
    "enable_early_termination": True,
    "termination_height": 0.3,
    "pose_termination": True,
    "pose_termination_dist": 0.5,
    "root_pos_termination_dist": 3.0,
    "root_rot_termination_angle": 0.8
}

Source: ig_parkour_env.py:967-982

DeepMimicEnv

DeepMimic-style motion tracking environment with motion library.

Initialization

from parc.motion_tracker.envs.ig_parkour.dm_env import DeepMimicEnv

dm_env = DeepMimicEnv(
    config=config,
    num_envs=1024,
    device="cuda:0",
    visualize=False,
    char_model=kinematic_model
)

Parameters:

config (dict): Configuration with DeepMimic settings
num_envs (int): Number of parallel environments
device (str): PyTorch device
visualize (bool): Enable visualization
char_model: Kinematic character model

Source: dm_env.py:19-93

Motion Library

# Motion library manages reference motions
dm_env._mlib.num_motions()           # Total number of motions
dm_env._mlib.sample_motions(n)       # Sample n motion IDs
dm_env._mlib.get_motion_length(ids)  # Get motion durations

Motion sampling with failure rates:

# Motions with higher failure rates get more samples
weights_modifier = torch.clamp(
    dm_env._motion_id_fail_rates,
    min=min_motion_weight
)
motion_weights = weights_modifier * dm_env._mlib._motion_weights
motion_ids = dm_env._mlib.sample_motions(n, motion_weights)

Source: dm_env.py:473-521

Terrain Management

build_terrain()

Builds procedural terrain for motions.

verts, tris = dm_env.build_terrain(
    env_config=config["env"],
    terrain_save_path="terrain.pkl",
    x_offset=0.0,
    y_offset=0.0
)

Terrain build modes:

"square": Grid layout of terrains
"wide": Linear arrangement
"file": Load from motion files

Source: dm_env.py:95-316

load_terrain()

Loads pre-built terrain from file.

verts, tris = dm_env.load_terrain(terrain_save_path)

Source: dm_env.py:447-463

Reference Motion Updates

_update_ref_motion()

Updates reference motion state for current timestep.

dm_env._update_ref_motion()

# Access reference state
root_pos = dm_env._ref_root_pos      # [num_envs, 3]
root_rot = dm_env._ref_root_rot      # [num_envs, 4]
joint_rot = dm_env._ref_joint_rot    # [num_envs, num_joints-1, 3]
contacts = dm_env._ref_contacts      # [num_envs, num_contact_bodies]

Source: dm_env.py:523-545

compute_tar_obs()

Computes target observations for future timesteps.

tar_root_pos, tar_root_rot, tar_joint_rot, tar_key_pos, tar_contacts = \
    dm_env.compute_tar_obs(
        tar_obs_steps=torch.tensor([1, 2, 3]),
        env_ids=None
    )

# Returns:
# tar_root_pos: [num_envs, num_steps, 3]
# tar_root_rot: [num_envs, num_steps, 4]
# tar_joint_rot: [num_envs, num_steps, num_joints-1, 3]
# tar_key_pos: [num_envs, num_steps, num_key_bodies, 3]
# tar_contacts: [num_envs, num_steps, num_contact_bodies]

Source: dm_env.py:594-626

Failure Rate Tracking

# Exponential moving average of failure rates per motion
dm_env._motion_id_fail_rates  # [num_motions]
dm_env._ema_weight = 0.01     # EMA update weight

# Updated on episode termination (dm_env.py:650-661)
if done_flag == FAIL:
    fail_rate[motion_id] = fail_rate[motion_id] * (1 - ema) + ema
elif done_flag == SUCC or TIME:
    fail_rate[motion_id] = fail_rate[motion_id] * (1 - ema)

Source: dm_env.py:628-665

Utility Methods

# Get motion information for specific environment
motion_name = dm_env.get_env_motion_name(env_id)
motion_length = dm_env.get_env_motion_length(env_id)
motion_time = dm_env.get_env_motion_time(env_id)

# Demo mode and motion control
dm_env.set_demo_mode(val=True)
dm_env.set_motion_start_time_fraction(fraction_tensor)

# Logging
extra_info = dm_env.get_extra_log_info()
# Returns failure rates and statistics

Source: dm_env.py:737-758

Configuration Example

env:
  # Basic settings
  num_envs: 4096
  episode_length: 10.0
  timestep: 0.0166  # 60 Hz
  
  # Termination
  enable_early_termination: true
  termination_height: 0.3
  pose_termination: true
  pose_termination_dist: 0.5
  
  # Observations
  enable_tar_obs: true
  tar_obs_steps: [1, 2, 3]
  use_contact_info: true
  global_root_height_obs: true
  
  # Rewards
  pose_w: 0.5
  vel_w: 0.05
  root_pos_w: 0.15
  root_vel_w: 0.1
  key_pos_w: 0.2
  
  # DeepMimic settings
  fraction_dm_envs: 1.0
  dm:
    motion_file: "motions/motion_lib.yaml"
    terrains_per_motion: 1
    terrain_build_mode: "square"
    fail_rate_quantiles: [0.25, 0.5, 0.75]
    heightmap:
      horizontal_scale: 0.05
      padding: 1.0

Motion Generation

Animation

Motion Synthesis

Motion Tracking

Scripts

Environments

IGParkourEnv

Initialization

Key Attributes

Observation Space

Action Space

Core Methods

reset()

step()

Reward Structure

Termination Conditions

DeepMimicEnv

Initialization

Motion Library

Terrain Management

build_terrain()

load_terrain()

Reference Motion Updates

_update_ref_motion()

compute_tar_obs()

Failure Rate Tracking

Utility Methods

Configuration Example

Build docs developers (and LLMs) love

Motion Generation

Animation

Motion Synthesis

Motion Tracking

Scripts

​IGParkourEnv

​Initialization

​Key Attributes

​Observation Space

​Action Space

​Core Methods

​reset()

​step()

​Reward Structure

​Termination Conditions

​DeepMimicEnv

​Initialization

​Motion Library

​Terrain Management

​build_terrain()

​load_terrain()

​Reference Motion Updates

​_update_ref_motion()

​compute_tar_obs()

​Failure Rate Tracking

​Utility Methods

​Configuration Example

Build docs developers (and LLMs) love

IGParkourEnv

Initialization

Key Attributes

Observation Space

Action Space

Core Methods

reset()

step()

Reward Structure

Termination Conditions

DeepMimicEnv

Initialization

Motion Library

Terrain Management

build_terrain()

load_terrain()

Reference Motion Updates

_update_ref_motion()

compute_tar_obs()

Failure Rate Tracking

Utility Methods

Configuration Example