[Question] Switching assets during reset

[guactaco12]

Hi there, I am trying to train a robot in an indoor environment setting, I have about a thousand different room layouts which I won't be able to run a thousand envs in parallel due to memory constraint. In this case, let's say I run 32 envs in parallel, is it possible for me to switch the room layout for each env during the reset so I can iterate through all the room layouts.

[guactaco12]

I dig through the docs more and I think the question becomes this: in _setup_scene in DirectRLEnv, is there a way I can iterate over num_envs to change the specific asset config in each environment?

[RandomOakForest]

Thank you for posting this. Note that _setup_scene is added to DirectRLEnv by the user. Currently, programmatic per-env asset modification requires run-time scripting after objects are spawned. This is a great topic for our Discussions section. I'll move this question there. Here are a few pointers to consider.

After environments are cloned with scene.clone_environments, the assets (e.g., robots, objects) are available as batched asset objects (like Articulation or RigidObject) inside the environment (typically as attributes or in scene.articulations/scene.rigid_objects). The API is designed for vectorization: methods and properties act on all environments as multi-instance tensors, but they accept per-environment indexing for modification. See docs for ref.

Key Points

• Each asset’s .data property is a "batched" container, and many set/write methods can take an env_ids argument to modify only specific environments.
• Change values (e.g., root pose, color, mass) by first editing the buffer—often a torch.Tensor with shape [num_envs, ...]—then calling a relevant .write_*_to_sim method with the right indices.

Example: Change Root Pose of Each Robot in Select Environments

Suppose there is an articulated robot stored in self.robot (an Articulation), and after cloning, one wants to set a custom root position for environments 1 and 2:

import torch

# Suppose self.robot is an Articulation after cloning. data.default_root_state is [num_envs, 13]
# (x, y, z, qw, qx, qy, qz, vx, vy, vz, wx, wy, wz)
# Let's move environments 1 and 2 along x and y axes:

env_ids = [1, 2]
new_positions = torch.tensor([
    [2.0, 0.0, 0.0],  # For env 1
    [0.0, 2.0, 0.0],  # For env 2
], device=self.robot.device)

# Get current root states and modify positions
root_states = self.robot.data.default_root_state[env_ids].clone()
root_states[:, :3] = new_positions

# Write only to envs 1 and 2
self.robot.write_root_state_to_sim(root_states, env_ids=env_ids)

This code modifies only the specified environments—other environments remain unchanged.

You can use similar tensor-indexed logic to adjust other properties, but you must check if the property supports at-runtime change (mass or color, for instance, might require different API calls or be limited by USD prim semantics).