source/isaaclab_mimic/test/test_selection_strategy.py

# Copyright (c) 2024-2026, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md).
# All rights reserved.
#
# SPDX-License-Identifier: Apache-2.0

from isaaclab.app import AppLauncher

# launch omniverse app
simulation_app = AppLauncher(headless=True).app

import numpy as np
import pytest
import torch

import isaaclab.utils.math as PoseUtils

from isaaclab_mimic.datagen.datagen_info import DatagenInfo

# Importing the necessary classes for the testing
from isaaclab_mimic.datagen.selection_strategy import (
    NearestNeighborObjectStrategy,
    NearestNeighborRobotDistanceStrategy,
)

# Number of iterations to run the batched tests
NUM_ITERS = 1000


@pytest.fixture
def nearest_neighbor_object_strategy():
    """Fixture for NearestNeighborObjectStrategy."""
    return NearestNeighborObjectStrategy()


@pytest.fixture
def nearest_neighbor_robot_distance_strategy():
    """Fixture for NearestNeighborRobotDistanceStrategy."""
    return NearestNeighborRobotDistanceStrategy()


def test_select_source_demo_identity_orientations_object_strategy(nearest_neighbor_object_strategy):
    """Test the selection of source demonstrations using two distinct object_pose clusters.

    This method generates two clusters of object poses and randomly adjusts the current object pose within
    specified deviations. It then simulates multiple selections to verify that when the current pose is close
    to cluster 1, all selected indices correspond to that cluster, and that the same holds true for cluster 2.
    """

    # Define ranges for two clusters of object poses
    cluster_1_range_min = 0
    cluster_1_range_max = 4
    cluster_2_range_min = 25
    cluster_2_range_max = 35

    # Generate object poses for cluster 1 with varying translations
    src_object_poses_in_world_cluster_1 = [
        torch.eye(4) + torch.tensor([[0.0, 0.0, 0.0, i], [0.0, 0.0, 0.0, i], [0.0, 0.0, 0.0, i], [0.0, 0.0, 0.0, -1.0]])
        for i in range(cluster_1_range_min, cluster_1_range_max)
    ]

    # Generate object poses for cluster 2 similarly
    src_object_poses_in_world_cluster_2 = [
        torch.eye(4) + torch.tensor([[0.0, 0.0, 0.0, i], [0.0, 0.0, 0.0, i], [0.0, 0.0, 0.0, i], [0.0, 0.0, 0.0, -1.0]])
        for i in range(cluster_2_range_min, cluster_2_range_max)
    ]

    # Combine the poses from both clusters into a single list
    src_object_poses_in_world = src_object_poses_in_world_cluster_1 + src_object_poses_in_world_cluster_2

    # Create DatagenInfo instances for these positions
    src_subtask_datagen_infos = [
        DatagenInfo(object_poses={0: object_pose.unsqueeze(0)}) for object_pose in src_object_poses_in_world
    ]

    # Define the end-effector pose (not used in the nearest neighbor selection)
    eef_pose = torch.eye(4)

    # Test 1:
    # Set the current object pose to the first value of cluster 1 and add some noise
    # Check that the nearest neighbor is always part of cluster 1
    max_deviation = 3  # Define a maximum deviation for the current pose
    # Randomly select an index from cluster 1
    random_index_cluster_1 = np.random.randint(0, len(src_object_poses_in_world_cluster_1))
    cluster_1_curr_object_pose = src_object_poses_in_world_cluster_1[
        random_index_cluster_1
    ].clone()  # Use clone to avoid reference issues
    # Randomly adjust the current pose within the maximum deviation
    cluster_1_curr_object_pose[0, 3] += torch.rand(1).item() * max_deviation
    cluster_1_curr_object_pose[1, 3] += torch.rand(1).item() * max_deviation
    cluster_1_curr_object_pose[2, 3] += torch.rand(1).item() * max_deviation

    # Select source demonstrations multiple times to check randomness
    selected_indices = [
        nearest_neighbor_object_strategy.select_source_demo(
            eef_pose,
            cluster_1_curr_object_pose,
            src_subtask_datagen_infos,
            pos_weight=1.0,
            rot_weight=1.0,
            nn_k=3,  # Check among the top 3 nearest neighbors
        )
        for _ in range(NUM_ITERS)
    ]

    # Assert that all selected indices are valid indices within cluster 1
    assert np.all(np.array(selected_indices) < len(src_object_poses_in_world_cluster_1)), (
        "Some selected indices are not part of cluster 1."
    )

    # Test 2:
    # Set the current object pose to the first value of cluster 2 and add some noise
    # Check that the nearest neighbor is always part of cluster 2
    max_deviation = 5  # Define a maximum deviation for the current pose in cluster 2
    # Randomly select an index from cluster 2
    random_index_cluster_2 = np.random.randint(0, len(src_object_poses_in_world_cluster_2))
    cluster_2_curr_object_pose = src_object_poses_in_world_cluster_2[
        random_index_cluster_2
    ].clone()  # Use clone to avoid reference issues
    # Randomly adjust the current pose within the maximum deviation
    cluster_2_curr_object_pose[0, 3] += torch.rand(1).item() * max_deviation
    cluster_2_curr_object_pose[1, 3] += torch.rand(1).item() * max_deviation
    cluster_2_curr_object_pose[2, 3] += torch.rand(1).item() * max_deviation

    # Select source demonstrations multiple times to check randomness
    selected_indices = [
        nearest_neighbor_object_strategy.select_source_demo(
            eef_pose,
            cluster_2_curr_object_pose,
            src_subtask_datagen_infos,
            pos_weight=1.0,
            rot_weight=1.0,
            nn_k=6,  # Check among the top 6 nearest neighbors
        )
        for _ in range(20)
    ]

    # Assert that all selected indices are valid indices within cluster 2
    assert np.all(np.array(selected_indices) < len(src_object_poses_in_world)), (
        "Some selected indices are not part of cluster 2."
    )
    assert np.all(np.array(selected_indices) > (len(src_object_poses_in_world_cluster_1) - 1)), (
        "Some selected indices are not part of cluster 2."
    )


def test_select_source_demo_identity_orientations_robot_distance_strategy(nearest_neighbor_robot_distance_strategy):
    """Test the selection of source demonstrations based on identity-oriented poses with varying positions.

    This method generates two clusters of object poses and randomly adjusts the current object pose within
    specified deviations. It then simulates multiple selections to verify that when the current pose is close
    to cluster 1, all selected indices correspond to that cluster, and that the same holds true for cluster 2.
    """

    # Define ranges for two clusters of object poses
    cluster_1_range_min = 0
    cluster_1_range_max = 4
    cluster_2_range_min = 25
    cluster_2_range_max = 35

    # Generate random transformed object poses for cluster 1 with varying translations
    # This represents the first object pose for the transformed subtask segment for each source demo
    transformed_eef_pose_cluster_1 = [
        torch.eye(4) + torch.tensor([[0, 0, 0, i], [0, 0, 0, i], [0, 0, 0, i], [0, 0, 0, -1]])
        for i in range(cluster_1_range_min, cluster_1_range_max)
    ]

    # Generate object poses for cluster 2 similarly
    transformed_eef_pose_cluster_2 = [
        torch.eye(4) + torch.tensor([[0, 0, 0, i], [0, 0, 0, i], [0, 0, 0, i], [0, 0, 0, -1]])
        for i in range(cluster_2_range_min, cluster_2_range_max)
    ]

    # Combine the poses from both clusters into a single list
    # This represents the first end effector pose for the transformed subtask segment for each source demo
    transformed_eef_in_world_poses_tensor = torch.stack(transformed_eef_pose_cluster_1 + transformed_eef_pose_cluster_2)

    # Create transformation matrices corresponding to each source object pose
    src_obj_in_world_poses = torch.stack(
        [
            PoseUtils.generate_random_transformation_matrix(pos_boundary=10, rot_boundary=(2 * np.pi))
            for _ in range(transformed_eef_in_world_poses_tensor.shape[0])
        ]
    )

    # Calculate the src_eef poses from the transformed eef poses, src_obj_in_world and curr_obj_pose_in_world
    # This is the inverse of the transformation of the eef pose done in NearestNeighborRobotDistanceStrategy
    # Refer to NearestNeighborRobotDistanceStrategy.select_source_demo for more details
    curr_object_in_world_pose = PoseUtils.generate_random_transformation_matrix(
        pos_boundary=10, rot_boundary=(2 * np.pi)
    )
    world_in_curr_obj_pose = PoseUtils.pose_inv(curr_object_in_world_pose)

    src_eef_in_src_obj_poses = PoseUtils.pose_in_A_to_pose_in_B(
        pose_in_A=transformed_eef_in_world_poses_tensor,
        pose_A_in_B=world_in_curr_obj_pose,
    )

    src_eef_in_world_poses = PoseUtils.pose_in_A_to_pose_in_B(
        pose_in_A=src_eef_in_src_obj_poses,
        pose_A_in_B=src_obj_in_world_poses,
    )

    # Check that both lists have the same length
    assert src_obj_in_world_poses.shape[0] == src_eef_in_world_poses.shape[0], (
        "Source object poses and end effector poses does not have the same length. "
        "This is a bug in the test code and not the source code."
    )

    # Create DatagenInfo instances for these positions
    src_subtask_datagen_infos = [
        DatagenInfo(eef_pose=src_eef_in_world_pose.unsqueeze(0), object_poses={0: src_obj_in_world_pose.unsqueeze(0)})
        for src_obj_in_world_pose, src_eef_in_world_pose in zip(src_obj_in_world_poses, src_eef_in_world_poses)
    ]

    # Test 1: Ensure the nearest neighbor is always part of cluster 1
    max_deviation = 3  # Define a maximum deviation for the current pose
    # Define the end-effector pose
    # Set the current object pose to the first value of cluster 1 and add some noise
    random_index_cluster_1 = np.random.randint(0, len(transformed_eef_pose_cluster_1))
    curr_eef_in_world_pose = transformed_eef_pose_cluster_1[
        random_index_cluster_1
    ].clone()  # Use clone to avoid reference issues
    # Randomly adjust the current pose within the maximum deviation
    curr_eef_in_world_pose[0, 3] += torch.rand(1).item() * max_deviation
    curr_eef_in_world_pose[1, 3] += torch.rand(1).item() * max_deviation
    curr_eef_in_world_pose[2, 3] += torch.rand(1).item() * max_deviation

    # Select source demonstrations multiple times to check randomness
    selected_indices = [
        nearest_neighbor_robot_distance_strategy.select_source_demo(
            curr_eef_in_world_pose,
            curr_object_in_world_pose,
            src_subtask_datagen_infos,
            pos_weight=1.0,
            rot_weight=1.0,
            nn_k=3,  # Check among the top 3 nearest neighbors
        )
        for _ in range(20)
    ]

    # Assert that all selected indices are valid indices within cluster 1
    assert np.all(np.array(selected_indices) < len(transformed_eef_pose_cluster_1)), (
        "Some selected indices are not part of cluster 1."
    )

    # Test 2: Ensure the nearest neighbor is always part of cluster 2
    max_deviation = 3  # Define a maximum deviation for the current pose
    # Define the end-effector pose
    # Set the current object pose to the first value of cluster 2 and add some noise
    random_index_cluster_2 = np.random.randint(0, len(transformed_eef_pose_cluster_2))
    curr_eef_in_world_pose = transformed_eef_pose_cluster_2[
        random_index_cluster_2
    ].clone()  # Use clone to avoid reference issues
    # Randomly adjust the current pose within the maximum deviation
    curr_eef_in_world_pose[0, 3] += torch.rand(1).item() * max_deviation
    curr_eef_in_world_pose[1, 3] += torch.rand(1).item() * max_deviation
    curr_eef_in_world_pose[2, 3] += torch.rand(1).item() * max_deviation

    # Select source demonstrations multiple times to check randomness
    selected_indices = [
        nearest_neighbor_robot_distance_strategy.select_source_demo(
            curr_eef_in_world_pose,
            curr_object_in_world_pose,
            src_subtask_datagen_infos,
            pos_weight=1.0,
            rot_weight=1.0,
            nn_k=3,  # Check among the top 3 nearest neighbors
        )
        for _ in range(20)
    ]

    # Assert that all selected indices are valid indices within cluster 2
    assert np.all(np.array(selected_indices) < transformed_eef_in_world_poses_tensor.shape[0]), (
        "Some selected indices are not part of cluster 2."
    )
    assert np.all(np.array(selected_indices) > (len(transformed_eef_pose_cluster_1) - 1)), (
        "Some selected indices are not part of cluster 2."
    )