robosuite/controllers/composite/composite_controller.py

import json
import re
from collections import OrderedDict
from typing import Dict, Optional, Tuple

import numpy as np

from robosuite.controllers import controller_factory
from robosuite.models.grippers.gripper_model import GripperModel
from robosuite.models.robots.robot_model import RobotModel
from robosuite.utils.binding_utils import MjSim
from robosuite.utils.ik_utils import IKSolver, get_nullspace_gains
from robosuite.utils.log_utils import ROBOSUITE_DEFAULT_LOGGER

REGISTERED_COMPOSITE_CONTROLLERS_DICT = {}


def register_composite_controller(target_class):
    if not hasattr(target_class, "name"):
        ROBOSUITE_DEFAULT_LOGGER.warning(
            "The name of the composite controller is not specified. Using the class name as the key."
        )
        key = "_".join(re.sub(r"([A-Z0-9])", r" \1", target_class.__name__).split()).upper()
    else:
        key = target_class.name
    REGISTERED_COMPOSITE_CONTROLLERS_DICT[key] = target_class
    return target_class


@register_composite_controller
class CompositeController:
    """This is the parent class for all composite controllers. If you want to develop an advanced version of your controller, you should subclass from this composite controller."""

    name = "BASIC"

    def __init__(self, sim: MjSim, robot_model: RobotModel, grippers: Dict[str, GripperModel]):
        # TODO: grippers repeat with members inside robot_model. Currently having this additioanl field to make naming query easy.
        self.sim = sim
        self.robot_model = robot_model

        self.grippers = grippers

        self.part_controllers = OrderedDict()

        self._action_split_indexes = OrderedDict()

        self.part_controller_config = None

        self.arms = self.robot_model.arms

        self._applied_action_dict = {}

    def load_controller_config(
        self, part_controller_config, composite_controller_specific_config: Optional[Dict] = None
    ):
        self.composite_controller_specific_config = composite_controller_specific_config
        body_part_ordering = self.composite_controller_specific_config.get("body_part_ordering", None)
        if body_part_ordering is not None:
            self.part_controller_config = OrderedDict()
            assert len(body_part_ordering) == len(part_controller_config)
            for part_name in body_part_ordering:
                self.part_controller_config[part_name] = part_controller_config[part_name]
        else:
            self.part_controller_config = part_controller_config

        self.part_controllers.clear()
        self._action_split_indexes.clear()
        self._init_controllers()
        self._validate_composite_controller_specific_config()
        self.setup_action_split_idx()

    def _init_controllers(self):
        for part_name in self.part_controller_config.keys():
            controller_params = self.part_controller_config[part_name]
            self.part_controllers[part_name] = controller_factory(
                part_name=part_name,
                controller_type=self.part_controller_config[part_name]["type"],
                controller_params=controller_params,
            )

    def _validate_composite_controller_specific_config(self):
        """Some aspects of composite controller specific configs, if they exist,
        may only be verified once the part controller are initialized."""
        pass

    def setup_action_split_idx(self):
        previous_idx = 0
        last_idx = 0
        for part_name, controller in self.part_controllers.items():
            if part_name in self.grippers.keys():
                last_idx += self.grippers[part_name].dof
            else:
                last_idx += controller.control_dim
            self._action_split_indexes[part_name] = (previous_idx, last_idx)
            previous_idx = last_idx

    def set_goal(self, all_action):
        for part_name, controller in self.part_controllers.items():
            start_idx, end_idx = self._action_split_indexes[part_name]
            action = all_action[start_idx:end_idx]
            if part_name in self.grippers.keys():
                action = self.grippers[part_name].format_action(action)
            controller.set_goal(action)

    def reset(self):
        for part_name, controller in self.part_controllers.items():
            controller.reset_goal()

    def run_controller(self, enabled_parts):
        self.update_state()
        self._applied_action_dict.clear()
        for part_name, controller in self.part_controllers.items():
            if enabled_parts.get(part_name, False):
                self._applied_action_dict[part_name] = controller.run_controller()

        return self._applied_action_dict

    def get_control_dim(self, part_name):
        if part_name not in self.part_controllers:
            return 0
        else:
            return self.part_controllers[part_name].control_dim

    def get_controller_base_pose(self, controller_name):
        """
        Get the base position and orientation of a specified controller's part. Note: this pose may likely differ from
        the robot base's pose.

        Args:
            controller_name (str): The name of the controller, used to look up part-specific information.

        Returns:
            tuple[np.ndarray, np.ndarray]: A tuple containing:
                - base_pos (np.ndarray): The 3D position of the part's center in world coordinates (shape: (3,)).
                - base_ori (np.ndarray): The 3x3 rotation matrix representing the part's orientation in world coordinates.

        Details:
            - Uses the controller's `naming_prefix` and `part_name` to construct the corresponding site name.
            - Queries the simulation (`self.sim`) for the site's position (`site_xpos`) and orientation (`site_xmat`).
            - The site orientation matrix is reshaped from a flat array of size 9 to a 3x3 rotation matrix.
        """
        naming_prefix = self.part_controllers[controller_name].naming_prefix
        part_name = self.part_controllers[controller_name].part_name
        base_pos = np.array(self.sim.data.site_xpos[self.sim.model.site_name2id(f"{naming_prefix}{part_name}_center")])
        base_ori = np.array(
            self.sim.data.site_xmat[self.sim.model.site_name2id(f"{naming_prefix}{part_name}_center")].reshape([3, 3])
        )
        return base_pos, base_ori

    def update_state(self):
        for arm in self.arms:
            base_pos, base_ori = self.get_controller_base_pose(controller_name=arm)
            self.part_controllers[arm].update_origin(base_pos, base_ori)

    def get_controller(self, part_name):
        return self.part_controllers[part_name]

    @property
    def action_limits(self):
        low, high = [], []
        for part_name, controller in self.part_controllers.items():
            if part_name not in self.arms:
                if part_name in self.grippers.keys():
                    low_g, high_g = (
                        [-1] * self.grippers[part_name].dof,
                        [1] * self.grippers[part_name].dof,
                    )
                    low, high = np.concatenate([low, low_g]), np.concatenate([high, high_g])
                else:
                    control_dim = controller.control_dim
                    low_c, high_c = ([-1] * control_dim, [1] * control_dim)
                    low, high = np.concatenate([low, low_c]), np.concatenate([high, high_c])
            else:
                low_c, high_c = controller.control_limits
                low, high = np.concatenate([low, low_c]), np.concatenate([high, high_c])
        return low, high

    def create_action_vector(self, action_dict):
        """
        A helper function that creates the action vector given a dictionary
        """
        full_action_vector = np.zeros(self.action_limits[0].shape)
        for part_name, action_vector in action_dict.items():
            if part_name not in self._action_split_indexes:
                ROBOSUITE_DEFAULT_LOGGER.debug(f"{part_name} is not specified in the action space")
                continue
            start_idx, end_idx = self._action_split_indexes[part_name]
            if end_idx - start_idx == 0:
                # skipping not controlling actions
                continue
            assert len(action_vector) == (end_idx - start_idx), ROBOSUITE_DEFAULT_LOGGER.error(
                f"Action vector for {part_name} is not the correct size. Expected {end_idx - start_idx} for {part_name}, got {len(action_vector)}"
            )
            full_action_vector[start_idx:end_idx] = action_vector
        return full_action_vector

    def create_action_dict_from_action_vector(self, action_vector):
        action_dict = {}
        for part_name, (start_idx, end_idx) in self._action_split_indexes.items():
            action_dict[part_name] = action_vector[start_idx:end_idx]
        return action_dict

    def get_action_info(self):
        action_index_info = []
        action_dim_info = []
        for part_name, (
            start_idx,
            end_idx,
        ) in self._action_split_indexes.items():
            action_dim_info.append(f"{part_name}: {(end_idx - start_idx)} dim")
            action_index_info.append(f"{part_name}: {start_idx}:{end_idx}")

        return action_index_info, action_dim_info

    def print_action_info(self):
        action_index_info, action_dim_info = self.get_action_info()
        action_index_info_str = ", ".join(action_index_info)
        action_dim_info_str = ", ".join(action_dim_info)
        ROBOSUITE_DEFAULT_LOGGER.info(f"Action Dimensions: [{action_dim_info_str}]")
        ROBOSUITE_DEFAULT_LOGGER.info(f"Action Indices: [{action_index_info_str}]")

    def get_action_info_dict(self):
        info_dict = {}
        info_dict["Action Dimension"] = self.action_limits[0].shape
        info_dict.update(dict(self._action_split_indexes))
        return info_dict

    def print_action_info_dict(self, name: str = ""):
        info_dict = self.get_action_info_dict()
        info_dict_str = f"\nAction Info for {name}:\n\n{json.dumps(dict(info_dict), indent=4)}"
        ROBOSUITE_DEFAULT_LOGGER.info(info_dict_str)


@register_composite_controller
class HybridMobileBase(CompositeController):
    name = "HYBRID_MOBILE_BASE"

    def set_goal(self, all_action):
        action_mode = all_action[-1]
        if action_mode > 0:
            update_wrt_origin = True
        else:
            update_wrt_origin = False

        for part_name, controller in self.part_controllers.items():
            start_idx, end_idx = self._action_split_indexes[part_name]
            action = all_action[start_idx:end_idx]
            if part_name in self.grippers.keys():
                action = self.grippers[part_name].format_action(action)

            if part_name in self.arms and hasattr(controller, "set_goal_update_mode"):
                """
                Query the last action dimension to determine if using
                base mode (value > 0) or arm mode (value < 1).
                If in base mode, update the new arm goal based on current desired goal,
                to have the arm tracking with the reset of the moving base
                as closely as possible without lagging or overshooting.
                """
                action_mode = all_action[-1]
                if action_mode > 0:
                    goal_update_mode = "desired"
                else:
                    goal_update_mode = "achieved"
                controller.set_goal_update_mode(goal_update_mode)

            controller.set_goal(action)

    @property
    def action_limits(self):
        low, high = super().action_limits
        return np.concatenate((low, [-1])), np.concatenate((high, [1]))

    def create_action_vector(self, action_dict):
        """
        A helper function that creates the action vector given a dictionary
        """
        full_action_vector = np.zeros(self.action_limits[0].shape)
        for part_name, action_vector in action_dict.items():
            if part_name not in self._action_split_indexes:
                ROBOSUITE_DEFAULT_LOGGER.debug(f"{part_name} is not specified in the action space")
                continue
            start_idx, end_idx = self._action_split_indexes[part_name]
            if end_idx - start_idx == 0:
                # skipping not controlling actions
                continue
            assert len(action_vector) == (end_idx - start_idx), ROBOSUITE_DEFAULT_LOGGER.error(
                f"Action vector for {part_name} is not the correct size. Expected {end_idx - start_idx} for {part_name}, got {len(action_vector)}"
            )
            full_action_vector[start_idx:end_idx] = action_vector
        full_action_vector[-1] = action_dict.get("base_mode", -1)
        return full_action_vector


@register_composite_controller
class WholeBody(CompositeController):
    name = "WHOLE_BODY_COMPOSITE"

    def __init__(self, sim: MjSim, robot_model: RobotModel, grippers: Dict[str, GripperModel]):
        super().__init__(sim, robot_model, grippers)

        self.joint_action_policy: IKSolver = None
        # TODO: handle different types of joint action policies; joint_action_policy maps
        # task space actions (such as end effector poses) to joint actions (such as joint angles or joint torques)

        self._whole_body_controller_action_split_indexes: OrderedDict = OrderedDict()

    def _init_controllers(self):
        for part_name in self.part_controller_config.keys():
            controller_params = self.part_controller_config[part_name]
            self.part_controllers[part_name] = controller_factory(
                part_name=part_name,
                controller_type=self.part_controller_config[part_name]["type"],
                controller_params=controller_params,
            )
        # for whole body composite controller, validate before init_joint_action_policy
        self._validate_composite_controller_specific_config()
        self._init_joint_action_policy()

    def _init_joint_action_policy(self):
        """Joint action policy initialization.

        Joint action policy converts input targets (such as end-effector poses, head poses) to joint actions
        (such as joint angles or joint torques).

        Examples of joint_action_policy could be an IK policy, a neural network policy, a model predictive controller, etc.
        """
        raise NotImplementedError("WholeBody CompositeController requires a joint action policy")

    def setup_action_split_idx(self):
        """
        Action split indices for the underlying factorized controllers.

        WholeBody controller takes in a different action space from the
        underlying factorized controllers for individual body parts.
        """
        previous_idx = 0
        last_idx = 0
        # add joint_action_policy related body parts' action split index first
        for part_name in self.composite_controller_specific_config["actuation_part_names"]:
            try:
                last_idx += self.part_controllers[part_name].control_dim
            except KeyError as e:
                import ipdb

                ipdb.set_trace()
                raise KeyError(f"Part name '{part_name}' not found in part_controllers: {e}")
            self._action_split_indexes[part_name] = (previous_idx, last_idx)
            previous_idx = last_idx

        for part_name, controller in self.part_controllers.items():
            if part_name not in self.composite_controller_specific_config["actuation_part_names"]:
                if part_name in self.grippers.keys():
                    last_idx += self.grippers[part_name].dof
                else:
                    last_idx += controller.control_dim
                self._action_split_indexes[part_name] = (previous_idx, last_idx)
                previous_idx = last_idx

        self.setup_whole_body_controller_action_split_idx()

    def setup_whole_body_controller_action_split_idx(self):
        """
        Action split indices for the composite controller's input action space.

        WholeBodyIK composite controller takes in a different action space from the
        underlying factorized controllers.
        """
        # add joint_action_policy's action split indexes first
        self._whole_body_controller_action_split_indexes.update(self.joint_action_policy.action_split_indexes())

        # prev and last index correspond to the IK solver indexes' last index
        previous_idx = last_idx = list(self._whole_body_controller_action_split_indexes.values())[-1][-1]
        for part_name, controller in self.part_controllers.items():
            if part_name in self.composite_controller_specific_config["actuation_part_names"]:
                continue
            if part_name in self.grippers.keys():
                last_idx += self.grippers[part_name].dof
            else:
                last_idx += controller.control_dim
            self._whole_body_controller_action_split_indexes[part_name] = (previous_idx, last_idx)
            previous_idx = last_idx

    def set_goal(self, all_action):
        if self.composite_controller_specific_config.get("skip_wbc_action", False):
            super().set_goal(all_action)
            return

        target_qpos = self.joint_action_policy.solve(all_action[: self.joint_action_policy.control_dim])
        # create new all_action vector with the IK solver's actions first
        all_action = np.concatenate([target_qpos, all_action[self.joint_action_policy.control_dim :]])
        for part_name, controller in self.part_controllers.items():
            start_idx, end_idx = self._action_split_indexes[part_name]
            action = all_action[start_idx:end_idx]
            if part_name in self.grippers.keys():
                action = self.grippers[part_name].format_action(action)
            controller.set_goal(action)

    def update_state(self):
        # no need for extra update state here, since Jacobians are computed inside the controllers of individual body parts
        return

    @property
    def action_limits(self):
        """
        Returns the action limits for the whole body controller.
        Corresponds to each term in the action vector passed to env.step().
        """
        if self.composite_controller_specific_config.get("skip_wbc_action", False):
            return super().action_limits

        low, high = [], []
        # assumption: IK solver's actions come first
        low_c, high_c = self.joint_action_policy.control_limits
        low, high = np.concatenate([low, low_c]), np.concatenate([high, high_c])
        for part_name, controller in self.part_controllers.items():
            # Exclude terms that the IK solver handles
            if part_name in self.composite_controller_specific_config["actuation_part_names"]:
                continue
            if part_name not in self.arms:
                if part_name in self.grippers.keys():
                    low_g, high_g = (
                        [-1] * self.grippers[part_name].dof,
                        [1] * self.grippers[part_name].dof,
                    )
                    low, high = np.concatenate([low, low_g]), np.concatenate([high, high_g])
                else:
                    control_dim = controller.control_dim
                    low_c, high_c = ([-1] * control_dim, [1] * control_dim)
                    low, high = np.concatenate([low, low_c]), np.concatenate([high, high_c])
            else:
                low_c, high_c = controller.control_limits
                low, high = np.concatenate([low, low_c]), np.concatenate([high, high_c])
        return low, high

    def create_action_vector(self, action_dict: Dict[str, np.ndarray]) -> np.ndarray:
        if self.composite_controller_specific_config.get("skip_wbc_action", False):
            return super().create_action_vector(action_dict)

        full_action_vector = np.zeros(self.action_limits[0].shape)
        for part_name, action_vector in action_dict.items():
            if part_name not in self._whole_body_controller_action_split_indexes:
                ROBOSUITE_DEFAULT_LOGGER.debug(f"{part_name} is not specified in the action space")
                continue
            start_idx, end_idx = self._whole_body_controller_action_split_indexes[part_name]
            if end_idx - start_idx == 0:
                # skipping not controlling actions
                continue
            assert len(action_vector) == (end_idx - start_idx), ROBOSUITE_DEFAULT_LOGGER.error(
                f"Action vector for {part_name} is not the correct size. Expected {end_idx - start_idx} for {part_name}, got {len(action_vector)}"
            )
            full_action_vector[start_idx:end_idx] = action_vector
        return full_action_vector

    def create_action_dict_from_action_vector(self, action_vector):
        if self.composite_controller_specific_config.get("skip_wbc_action", False):
            return super().create_action_dict_from_action_vector(action_vector)

        action_dict = {}
        for part_name, (start_idx, end_idx) in self._whole_body_controller_action_split_indexes.items():
            action_dict[part_name] = action_vector[start_idx:end_idx]
        return action_dict

    def get_action_info(self):
        if self.composite_controller_specific_config.get("skip_wbc_action", False):
            return super().get_action_info()

        action_index_info = []
        action_dim_info = []
        for part_name, (
            start_idx,
            end_idx,
        ) in self._whole_body_controller_action_split_indexes.items():
            action_dim_info.append(f"{part_name}: {(end_idx - start_idx)} dim")
            action_index_info.append(f"{part_name}: {start_idx}:{end_idx}")
        return action_index_info, action_dim_info

    def print_action_info(self):
        if self.composite_controller_specific_config.get("skip_wbc_action", False):
            return super().print_action_info()

        action_index_info, action_dim_info = self.get_action_info()
        action_index_info_str = ", ".join(action_index_info)
        action_dim_info_str = ", ".join(action_dim_info)
        ROBOSUITE_DEFAULT_LOGGER.info(f"Action Dimensions: [{action_dim_info_str}]")
        ROBOSUITE_DEFAULT_LOGGER.info(f"Action Indices: [{action_index_info_str}]")

    def get_action_info_dict(self):
        if self.composite_controller_specific_config.get("skip_wbc_action", False):
            return super().get_action_info_dict()

        info_dict = {}
        info_dict["Action Dimension"] = self.action_limits[0].shape
        info_dict.update(dict(self._whole_body_controller_action_split_indexes))
        return info_dict

    def get_action_info_dict(self):
        if self.composite_controller_specific_config.get("skip_wbc_action", False):
            return super().get_action_info_dict()

        info_dict = {}
        info_dict["Action Dimension"] = self.action_limits[0].shape
        info_dict.update(dict(self._whole_body_controller_action_split_indexes))
        return info_dict

    def print_action_info_dict(self, name: str = ""):
        if self.composite_controller_specific_config.get("skip_wbc_action", False):
            return super().print_action_info_dict(name)

        info_dict = self.get_action_info_dict()
        info_dict_str = f"\nAction Info for {name}:\n\n{json.dumps(dict(info_dict), indent=4)}"
        ROBOSUITE_DEFAULT_LOGGER.info(info_dict_str)


@register_composite_controller
class WholeBodyIK(WholeBody):
    name = "WHOLE_BODY_IK"

    def __init__(self, sim: MjSim, robot_model: RobotModel, grippers: Dict[str, GripperModel]):
        super().__init__(sim, robot_model, grippers)

    def _validate_composite_controller_specific_config(self) -> None:
        # Check that all actuation_part_names exist in part_controllers
        original_ik_controlled_parts = self.composite_controller_specific_config["actuation_part_names"]
        self.valid_ik_controlled_parts = []
        valid_ref_names = []

        assert (
            "ref_name" in self.composite_controller_specific_config
        ), "The 'ref_name' key is missing from composite_controller_specific_config."

        for part in original_ik_controlled_parts:
            if part in self.part_controllers:
                self.valid_ik_controlled_parts.append(part)
            else:
                ROBOSUITE_DEFAULT_LOGGER.warning(
                    f"Part '{part}' specified in 'actuation_part_names' "
                    "does not exist in part_controllers. Removing ..."
                )

        # Update the configuration with only the valid parts
        self.composite_controller_specific_config["actuation_part_names"] = self.valid_ik_controlled_parts

        # Loop through ref_names and validate against mujoco model
        original_ref_names = self.composite_controller_specific_config.get("ref_name", [])
        for ref_name in original_ref_names:
            if ref_name in self.sim.model.site_names:  # Check if the site exists in the mujoco model
                valid_ref_names.append(ref_name)
            else:
                ROBOSUITE_DEFAULT_LOGGER.warning(
                    f"Reference name '{ref_name}' specified in configuration"
                    " does not exist in the mujoco model. Removing ..."
                )

        # Update the configuration with only the valid reference names
        self.composite_controller_specific_config["ref_name"] = valid_ref_names

    def _init_joint_action_policy(self):
        joint_names: str = []
        for part_name in self.composite_controller_specific_config["actuation_part_names"]:
            if part_name in self.part_controllers:
                joint_names += self.part_controllers[part_name].joint_names
            else:
                ROBOSUITE_DEFAULT_LOGGER.warning(
                    f"{part_name} is not a valid part name in part_controllers." " Skipping this part for IK control."
                )

        # Compute nullspace gains, Kn.
        Kn = get_nullspace_gains(joint_names, self.composite_controller_specific_config["nullspace_joint_weights"])
        mocap_bodies = []
        robot_config = {
            "end_effector_sites": self.composite_controller_specific_config["ref_name"],
            "joint_names": joint_names,
            "mocap_bodies": mocap_bodies,
            "nullspace_gains": Kn,
        }
        self.joint_action_policy = IKSolver(
            model=self.sim.model._model,
            data=self.sim.data._data,
            robot_config=robot_config,
            damping=self.composite_controller_specific_config.get("ik_pseudo_inverse_damping", 5e-2),
            integration_dt=self.composite_controller_specific_config.get("ik_integration_dt", 0.1),
            max_dq=self.composite_controller_specific_config.get("ik_max_dq", 4),
            max_dq_torso=self.composite_controller_specific_config.get("ik_max_dq_torso", 0.2),
            input_rotation_repr=self.composite_controller_specific_config.get("ik_input_rotation_repr", "axis_angle"),
            input_ref_frame=self.composite_controller_specific_config.get("ik_input_ref_frame", "world"),
            input_type=self.composite_controller_specific_config.get("ik_input_type", "axis_angle"),
            debug=self.composite_controller_specific_config.get("verbose", False),
        )