🎲 Domain Randomization#

Randomize physics and visual properties for robust RL training.

Physics Randomization#

Apply after scene.build():

import genesis as gs
import torch

scene.build(n_envs=64)

# Friction randomization
robot.set_friction_ratio(
    friction_ratio=0.5 + torch.rand(scene.n_envs, robot.n_links),
    links_idx_local=range(robot.n_links),
)

# Mass randomization
robot.set_mass_shift(
    mass_shift=-0.5 + torch.rand(scene.n_envs, robot.n_links),
    links_idx_local=range(robot.n_links),
)

# Center of mass randomization
robot.set_COM_shift(
    com_shift=-0.05 + 0.1 * torch.rand(scene.n_envs, robot.n_links, 3),
    links_idx_local=range(robot.n_links),
)

Control Parameter Randomization#

import numpy as np

# Per-environment stiffness (Kp)
kp_values = 4000 + 1000 * np.random.rand(scene.n_envs, robot.n_dofs)
robot.set_dofs_kp(kp_values, motors_dof)

# Per-environment damping (Kv)
kv_values = 400 + 100 * np.random.rand(scene.n_envs, robot.n_dofs)
robot.set_dofs_kv(kv_values, motors_dof)

Object Position Randomization (Per-Episode)#

def reset_idx(self, envs_idx):
    num_reset = len(envs_idx)

    # Random object position
    random_x = torch.rand(num_reset, device=gs.device) * 0.4 + 0.2
    random_y = (torch.rand(num_reset, device=gs.device) - 0.5) * 0.5
    random_z = torch.ones(num_reset, device=gs.device) * 0.025
    random_pos = torch.stack([random_x, random_y, random_z], dim=-1)

    self.object.set_pos(random_pos, envs_idx=envs_idx)

Command Randomization#

def gs_rand(lower, upper, shape):
    """Uniform random in [lower, upper]"""
    return (upper - lower) * torch.rand(shape, device=gs.device) + lower

# Randomize velocity commands
commands = gs_rand(
    lower=torch.tensor([-1.0, -0.5, -0.5]),
    upper=torch.tensor([1.0, 0.5, 0.5]),
    shape=(self.num_envs, 3),
)

Available Methods#

Method

Shape

Description

set_friction_ratio

(n_envs, n_links)

Friction scaling

set_mass_shift

(n_envs, n_links)

Mass offset

set_COM_shift

(n_envs, n_links, 3)

COM offset

set_dofs_kp

(n_envs, n_dofs)

Position gain

set_dofs_kv

(n_envs, n_dofs)

Velocity gain

set_dofs_armature

(n_envs, n_dofs)

Motor inertia

Required Options#

Enable batching for physics randomization:

scene = gs.Scene(
    rigid_options=gs.options.RigidOptions(
        batch_dofs_info=True,
        batch_links_info=True,
    ),
)

Best Practices#

  1. Apply physics DR once after scene.build()

  2. Apply position/command DR at each episode reset

  3. Use envs_idx parameter for selective randomization

  4. Ensure tensor shapes match (n_envs, ...)

What’s Next#

  • Parallel Simulation - running batched environments for RL training

  • Locomotion - training a locomotion policy end-to-end with domain randomization