Environment config

Step 1 – Define the Standing Environment¶

In this step, you define what the robot’s world looks like and what it is asked to do.
Rather than building everything from scratch, you create a new environment configuration for the G1 standing task by subclassing Isaac Lab’s built-in:

isaaclab_tasks.manager_based.locomotion.velocity.config.g1.flat_env_cfg.G1FlatEnvCfg

We reuse G1FlatEnvCfg because it already knows how to:

Spawn the G1 robot on flat terrain
Set up sensors and basic locomotion rewards

Our job is to turn that locomotion task into a pure standing task by changing commands and rewards.

The new configs you will create are:

G1StandFlatEnvCfg – main training configuration
G1StandFlatPlayEnvCfg – play/visualization configuration

Both live in the g1_stand extension, not in Isaac Lab core.

Steps¶

Create and paste env config (g1_stand_env_cfg.py)

Create or edit this file: source/g1_stand/g1_stand/tasks/manager_based/g1_stand/g1_stand_env_cfg.py
Paste the code from the block below into that file (replace any existing template code).
Verify (optional): in a Python REPL with Isaac Lab, run: from g1_stand.g1_stand.tasks.manager_based.g1_stand.g1_stand_env_cfg import G1StandFlatEnvCfg and confirm it imports without errors.

from isaaclab.utils import configclass

from isaaclab_tasks.manager_based.locomotion.velocity.config.g1.flat_env_cfg import (
    G1FlatEnvCfg,
)


@configclass
class G1StandFlatEnvCfg(G1FlatEnvCfg):
    """G1 standing task on flat terrain (training config)."""

    def __post_init__(self) -> None:
        # Run parent post-init to build the standard G1 flat locomotion task
        super().__post_init__()

        # ------------------------------------------------------------------
        # Commands: force commanded base velocities to zero (stand still)
        # ------------------------------------------------------------------
        self.commands.base_velocity.ranges.lin_vel_x = (0.0, 0.0)
        self.commands.base_velocity.ranges.lin_vel_y = (0.0, 0.0)
        self.commands.base_velocity.ranges.ang_vel_z = (0.0, 0.0)

        # ------------------------------------------------------------------
        # Disable velocity tracking rewards (we are not tracking a velocity)
        # ------------------------------------------------------------------
        # These are defined in the G1 locomotion rewards config
        self.rewards.track_lin_vel_xy_exp.weight = 0.0
        self.rewards.track_ang_vel_z_exp.weight = 0.0

        # ------------------------------------------------------------------
        # Standing rewards
        # (weights are a reasonable starting point; you can tune later)
        # ------------------------------------------------------------------
        # Penalize terminations (falling, etc.)
        if hasattr(self.rewards, "termination_penalty"):
            self.rewards.termination_penalty.weight = -200.0

        # Keep the base upright on flat terrain
        if hasattr(self.rewards, "flat_orientation_l2"):
            self.rewards.flat_orientation_l2.weight = -1.0

        # Encourage a comfortable standing base height (only if defined)
        if hasattr(self.rewards, "base_height_l2"):
            self.rewards.base_height_l2.weight = -1.0

        # Discourage vertical bouncing
        if hasattr(self.rewards, "lin_vel_z_l2"):
            self.rewards.lin_vel_z_l2.weight = -0.5

        # Discourage sideways / forward angular motion of the base
        if hasattr(self.rewards, "ang_vel_xy_l2"):
            self.rewards.ang_vel_xy_l2.weight = -0.5

        # Keep joints close to a nominal standing pose (only if defined)
        if hasattr(self.rewards, "stand_still_joint_deviation_l1"):
            self.rewards.stand_still_joint_deviation_l1.weight = -0.2

        # Smooth actions
        if hasattr(self.rewards, "action_rate_l2"):
            self.rewards.action_rate_l2.weight = -0.01

        # Penalize large joint torques
        if hasattr(self.rewards, "dof_torques_l2"):
            self.rewards.dof_torques_l2.weight = -2.0e-6

        # ------------------------------------------------------------------
        # Episode / viewer tweaks for standing
        # ------------------------------------------------------------------
        # Shorter episodes are usually fine for a simple stand task
        self.episode_length_s = 10.0

        # Use a relatively large number of parallel envs for training
        # (you can downscale if your GPU is smaller)
        self.scene.num_envs = 4096
        self.scene.env_spacing = 3.0


@configclass
class G1StandFlatPlayEnvCfg(G1StandFlatEnvCfg):
    """G1 standing task on flat terrain (play/visualization config)."""

    def __post_init__(self) -> None:
        # Start from the training config settings
        super().__post_init__()

        # Fewer envs and more comfortable spacing for visualization
        self.scene.num_envs = 1
        self.scene.env_spacing = 3.0

        # Longer episodes for manual observation
        self.episode_length_s = 20.0

        # Disable observation corruption for cleaner visuals
        if hasattr(self.observations, "policy"):
            self.observations.policy.enable_corruption = False

        # Remove random pushes if they exist in the parent config
        if hasattr(self.events, "base_external_force_torque"):
            self.events.base_external_force_torque = None
        if hasattr(self.events, "push_robot"):
            self.events.push_robot = None

What this does¶

Inherits G1 flat locomotion
You get the full G1 robot setup, terrain, sensors, and basic reward structure from G1FlatEnvCfg.
Forces standing commands
Setting:

self.commands.base_velocity.ranges.lin_vel_x = (0.0, 0.0)
self.commands.base_velocity.ranges.lin_vel_y = (0.0, 0.0)
self.commands.base_velocity.ranges.ang_vel_z = (0.0, 0.0)

tells the command generator to always request zero velocity, i.e., stand still.

Disables velocity tracking rewards
With:

self.rewards.track_lin_vel_xy_exp.weight = 0.0
self.rewards.track_ang_vel_z_exp.weight = 0.0

you remove incentives to move at a nonzero commanded velocity.

Emphasizes standing-still rewards
We adjust weights for:
termination_penalty – strongly penalizes falling/termination.
flat_orientation_l2 – encourages upright torso.
base_height_l2 (if available) – encourages a desired base height.
lin_vel_z_l2 – discourages vertical bouncing.
ang_vel_xy_l2 – discourages tilting or spinning in roll/pitch.
stand_still_joint_deviation_l1 (if available) – keeps joints near nominal pose.
action_rate_l2 – smooths actions over time.
dof_torques_l2 – discourages large torques.

All of these are wrapped in hasattr checks so the code works even if some terms are not present in your Isaac Lab version.

Common mistake – missing reward attributes

Different Isaac Lab versions may not define all reward terms (for example, base_height_l2).
By checking with hasattr(self.rewards, "...") before setting weights, you avoid AttributeError at runtime.

Checklist¶

After editing g1_stand_env_cfg.py, confirm:

The file defines both G1StandFlatEnvCfg and G1StandFlatPlayEnvCfg.
The import in step 3 above runs without errors.
Commands are "stand still"; rewards prefer upright, low-motion, low-torque behavior.

When you run training (Step 4), the trainer will use this env config.