shadow_hand.py

# Copyright (c) 2018-2023, NVIDIA Corporation
# All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice, this
#    list of conditions and the following disclaimer.
#
# 2. Redistributions in binary form must reproduce the above copyright notice,
#    this list of conditions and the following disclaimer in the documentation
#    and/or other materials provided with the distribution.
#
# 3. Neither the name of the copyright holder nor the names of its
#    contributors may be used to endorse or promote products derived from
#    this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

import numpy as np
import os
import torch

from isaacgym import gymtorch
from isaacgym import gymapi

from isaacgymenvs.utils.torch_jit_utils import scale, unscale, quat_mul, quat_conjugate, quat_from_angle_axis, \
    to_torch, get_axis_params, torch_rand_float, tensor_clamp  

from isaacgymenvs.tasks.base.vec_task import VecTask


class ShadowHand(VecTask):

    def __init__(self, cfg, rl_device, sim_device, graphics_device_id, headless, virtual_screen_capture, force_render):

        self.cfg = cfg

        self.randomize = self.cfg["task"]["randomize"]
        self.randomization_params = self.cfg["task"]["randomization_params"]
        self.aggregate_mode = self.cfg["env"]["aggregateMode"]

        self.dist_reward_scale = self.cfg["env"]["distRewardScale"]
        self.rot_reward_scale = self.cfg["env"]["rotRewardScale"]
        self.action_penalty_scale = self.cfg["env"]["actionPenaltyScale"]
        self.success_tolerance = self.cfg["env"]["successTolerance"]
        self.reach_goal_bonus = self.cfg["env"]["reachGoalBonus"]
        self.fall_dist = self.cfg["env"]["fallDistance"]
        self.fall_penalty = self.cfg["env"]["fallPenalty"]
        self.rot_eps = self.cfg["env"]["rotEps"]

        self.vel_obs_scale = 0.2  # scale factor of velocity based observations
        self.force_torque_obs_scale = 10.0  # scale factor of velocity based observations

        self.reset_position_noise = self.cfg["env"]["resetPositionNoise"]
        self.reset_rotation_noise = self.cfg["env"]["resetRotationNoise"]
        self.reset_dof_pos_noise = self.cfg["env"]["resetDofPosRandomInterval"]
        self.reset_dof_vel_noise = self.cfg["env"]["resetDofVelRandomInterval"]

        self.force_scale = self.cfg["env"].get("forceScale", 0.0)
        self.force_prob_range = self.cfg["env"].get("forceProbRange", [0.001, 0.1])
        self.force_decay = self.cfg["env"].get("forceDecay", 0.99)
        self.force_decay_interval = self.cfg["env"].get("forceDecayInterval", 0.08)

        self.shadow_hand_dof_speed_scale = self.cfg["env"]["dofSpeedScale"]
        self.use_relative_control = self.cfg["env"]["useRelativeControl"]
        self.act_moving_average = self.cfg["env"]["actionsMovingAverage"]

        self.debug_viz = self.cfg["env"]["enableDebugVis"]

        self.max_episode_length = self.cfg["env"]["episodeLength"]
        self.reset_time = self.cfg["env"].get("resetTime", -1.0)
        self.print_success_stat = self.cfg["env"]["printNumSuccesses"]
        self.max_consecutive_successes = self.cfg["env"]["maxConsecutiveSuccesses"]
        self.av_factor = self.cfg["env"].get("averFactor", 0.1)

        self.object_type = self.cfg["env"]["objectType"]
        assert self.object_type in ["block", "egg", "pen"]

        self.ignore_z = (self.object_type == "pen")

        self.asset_files_dict = {
            "block": "urdf/objects/cube_multicolor.urdf",
            "egg": "mjcf/open_ai_assets/hand/egg.xml",
            "pen": "mjcf/open_ai_assets/hand/pen.xml"
        }

        if "asset" in self.cfg["env"]:
            self.asset_files_dict["block"] = self.cfg["env"]["asset"].get("assetFileNameBlock", self.asset_files_dict["block"])
            self.asset_files_dict["egg"] = self.cfg["env"]["asset"].get("assetFileNameEgg", self.asset_files_dict["egg"])
            self.asset_files_dict["pen"] = self.cfg["env"]["asset"].get("assetFileNamePen", self.asset_files_dict["pen"])

        # can be "openai", "full_no_vel", "full", "full_state"
        self.obs_type = self.cfg["env"]["observationType"]

        if not (self.obs_type in ["openai", "full_no_vel", "full", "full_state"]):
            raise Exception(
                "Unknown type of observations!\nobservationType should be one of: [openai, full_no_vel, full, full_state]")

        print("Obs type:", self.obs_type)

        self.num_obs_dict = {
            "openai": 42,
            "full_no_vel": 77,
            "full": 157,
            "full_state": 211
        }

        self.up_axis = 'z'

        self.fingertips = ["robot0:ffdistal", "robot0:mfdistal", "robot0:rfdistal", "robot0:lfdistal", "robot0:thdistal"]
        self.num_fingertips = len(self.fingertips)

        self.use_vel_obs = False
        self.fingertip_obs = True
        self.asymmetric_obs = self.cfg["env"]["asymmetric_observations"]

        num_states = 0
        if self.asymmetric_obs:
            num_states = 211

        self.cfg["env"]["numObservations"] = self.num_obs_dict[self.obs_type]
        self.cfg["env"]["numStates"] = num_states
        self.cfg["env"]["numActions"] = 20

        super().__init__(config=self.cfg, rl_device=rl_device, sim_device=sim_device, graphics_device_id=graphics_device_id, headless=headless, virtual_screen_capture=virtual_screen_capture, force_render=force_render)

        self.dt = self.sim_params.dt
        control_freq_inv = self.cfg["env"].get("controlFrequencyInv", 1)
        if self.reset_time > 0.0:
            self.max_episode_length = int(round(self.reset_time/(control_freq_inv * self.dt)))
            print("Reset time: ", self.reset_time)
            print("New episode length: ", self.max_episode_length)

        if self.viewer != None:
            cam_pos = gymapi.Vec3(10.0, 5.0, 1.0)
            cam_target = gymapi.Vec3(6.0, 5.0, 0.0)
            self.gym.viewer_camera_look_at(self.viewer, None, cam_pos, cam_target)

        # get gym GPU state tensors
        actor_root_state_tensor = self.gym.acquire_actor_root_state_tensor(self.sim)
        dof_state_tensor = self.gym.acquire_dof_state_tensor(self.sim)
        rigid_body_tensor = self.gym.acquire_rigid_body_state_tensor(self.sim)

        if self.obs_type == "full_state" or self.asymmetric_obs:
            sensor_tensor = self.gym.acquire_force_sensor_tensor(self.sim)
            self.vec_sensor_tensor = gymtorch.wrap_tensor(sensor_tensor).view(self.num_envs, self.num_fingertips * 6)

            dof_force_tensor = self.gym.acquire_dof_force_tensor(self.sim)
            self.dof_force_tensor = gymtorch.wrap_tensor(dof_force_tensor).view(self.num_envs, self.num_shadow_hand_dofs)

        self.gym.refresh_actor_root_state_tensor(self.sim)
        self.gym.refresh_dof_state_tensor(self.sim)
        self.gym.refresh_rigid_body_state_tensor(self.sim)

        # create some wrapper tensors for different slices
        self.shadow_hand_default_dof_pos = torch.zeros(self.num_shadow_hand_dofs, dtype=torch.float, device=self.device)
        self.dof_state = gymtorch.wrap_tensor(dof_state_tensor)
        self.shadow_hand_dof_state = self.dof_state.view(self.num_envs, -1, 2)[:, :self.num_shadow_hand_dofs]
        self.shadow_hand_dof_pos = self.shadow_hand_dof_state[..., 0]
        self.shadow_hand_dof_vel = self.shadow_hand_dof_state[..., 1]

        self.rigid_body_states = gymtorch.wrap_tensor(rigid_body_tensor).view(self.num_envs, -1, 13)
        self.num_bodies = self.rigid_body_states.shape[1]

        self.root_state_tensor = gymtorch.wrap_tensor(actor_root_state_tensor).view(-1, 13)

        self.num_dofs = self.gym.get_sim_dof_count(self.sim) // self.num_envs
        self.prev_targets = torch.zeros((self.num_envs, self.num_dofs), dtype=torch.float, device=self.device)
        self.cur_targets = torch.zeros((self.num_envs, self.num_dofs), dtype=torch.float, device=self.device)

        self.global_indices = torch.arange(self.num_envs * 3, dtype=torch.int32, device=self.device).view(self.num_envs, -1)
        self.x_unit_tensor = to_torch([1, 0, 0], dtype=torch.float, device=self.device).repeat((self.num_envs, 1))
        self.y_unit_tensor = to_torch([0, 1, 0], dtype=torch.float, device=self.device).repeat((self.num_envs, 1))
        self.z_unit_tensor = to_torch([0, 0, 1], dtype=torch.float, device=self.device).repeat((self.num_envs, 1))

        self.reset_goal_buf = self.reset_buf.clone()
        self.successes = torch.zeros(self.num_envs, dtype=torch.float, device=self.device)
        self.consecutive_successes = torch.zeros(1, dtype=torch.float, device=self.device)

        self.av_factor = to_torch(self.av_factor, dtype=torch.float, device=self.device)

        self.total_successes = 0
        self.total_resets = 0

        # object apply random forces parameters
        self.force_decay = to_torch(self.force_decay, dtype=torch.float, device=self.device)
        self.force_prob_range = to_torch(self.force_prob_range, dtype=torch.float, device=self.device)
        self.random_force_prob = torch.exp((torch.log(self.force_prob_range[0]) - torch.log(self.force_prob_range[1]))
                                           * torch.rand(self.num_envs, device=self.device) + torch.log(self.force_prob_range[1]))

        self.rb_forces = torch.zeros((self.num_envs, self.num_bodies, 3), dtype=torch.float, device=self.device)

    def create_sim(self):
        self.dt = self.cfg["sim"]["dt"]
        self.up_axis_idx = 2 if self.up_axis == 'z' else 1 # index of up axis: Y=1, Z=2

        self.sim = super().create_sim(self.device_id, self.graphics_device_id, self.physics_engine, self.sim_params)
        self._create_ground_plane()
        self._create_envs(self.num_envs, self.cfg["env"]['envSpacing'], int(np.sqrt(self.num_envs)))

        # If randomizing, apply once immediately on startup before the fist sim step
        if self.randomize:
            self.apply_randomizations(self.randomization_params)

    def _create_ground_plane(self):
        plane_params = gymapi.PlaneParams()
        plane_params.normal = gymapi.Vec3(0.0, 0.0, 1.0)
        self.gym.add_ground(self.sim, plane_params)

    def _create_envs(self, num_envs, spacing, num_per_row):
        lower = gymapi.Vec3(-spacing, -spacing, 0.0)
        upper = gymapi.Vec3(spacing, spacing, spacing)

        asset_root = os.path.normpath(os.path.join(os.path.dirname(os.path.abspath(__file__)), '../../assets'))
        shadow_hand_asset_file = os.path.normpath("mjcf/open_ai_assets/hand/shadow_hand.xml")

        if "asset" in self.cfg["env"]:
            # asset_root = self.cfg["env"]["asset"].get("assetRoot", asset_root)
            shadow_hand_asset_file = os.path.normpath(self.cfg["env"]["asset"].get("assetFileName", shadow_hand_asset_file))

        object_asset_file = self.asset_files_dict[self.object_type]

        # load shadow hand_ asset
        asset_options = gymapi.AssetOptions()
        asset_options.flip_visual_attachments = False
        asset_options.fix_base_link = True
        asset_options.collapse_fixed_joints = True
        asset_options.disable_gravity = True
        asset_options.thickness = 0.001
        asset_options.angular_damping = 0.01

        if self.physics_engine == gymapi.SIM_PHYSX:
            asset_options.use_physx_armature = True
        # Note - DOF mode is set in the MJCF file and loaded by Isaac Gym
        asset_options.default_dof_drive_mode = gymapi.DOF_MODE_NONE

        shadow_hand_asset = self.gym.load_asset(self.sim, asset_root, shadow_hand_asset_file, asset_options)

        self.num_shadow_hand_bodies = self.gym.get_asset_rigid_body_count(shadow_hand_asset)
        self.num_shadow_hand_shapes = self.gym.get_asset_rigid_shape_count(shadow_hand_asset)
        self.num_shadow_hand_dofs = self.gym.get_asset_dof_count(shadow_hand_asset)
        self.num_shadow_hand_actuators = self.gym.get_asset_actuator_count(shadow_hand_asset)
        self.num_shadow_hand_tendons = self.gym.get_asset_tendon_count(shadow_hand_asset)

        # tendon set up
        limit_stiffness = 30
        t_damping = 0.1
        relevant_tendons = ["robot0:T_FFJ1c", "robot0:T_MFJ1c", "robot0:T_RFJ1c", "robot0:T_LFJ1c"]
        tendon_props = self.gym.get_asset_tendon_properties(shadow_hand_asset)

        for i in range(self.num_shadow_hand_tendons):
            for rt in relevant_tendons:
                if self.gym.get_asset_tendon_name(shadow_hand_asset, i) == rt:
                    tendon_props[i].limit_stiffness = limit_stiffness
                    tendon_props[i].damping = t_damping
        self.gym.set_asset_tendon_properties(shadow_hand_asset, tendon_props)

        actuated_dof_names = [self.gym.get_asset_actuator_joint_name(shadow_hand_asset, i) for i in range(self.num_shadow_hand_actuators)]
        self.actuated_dof_indices = [self.gym.find_asset_dof_index(shadow_hand_asset, name) for name in actuated_dof_names]

        # get shadow_hand dof properties, loaded by Isaac Gym from the MJCF file
        shadow_hand_dof_props = self.gym.get_asset_dof_properties(shadow_hand_asset)

        self.shadow_hand_dof_lower_limits = []
        self.shadow_hand_dof_upper_limits = []
        self.shadow_hand_dof_default_pos = []
        self.shadow_hand_dof_default_vel = []

        for i in range(self.num_shadow_hand_dofs):
            self.shadow_hand_dof_lower_limits.append(shadow_hand_dof_props['lower'][i])
            self.shadow_hand_dof_upper_limits.append(shadow_hand_dof_props['upper'][i])
            self.shadow_hand_dof_default_pos.append(0.0)
            self.shadow_hand_dof_default_vel.append(0.0)

        self.actuated_dof_indices = to_torch(self.actuated_dof_indices, dtype=torch.long, device=self.device)
        self.shadow_hand_dof_lower_limits = to_torch(self.shadow_hand_dof_lower_limits, device=self.device)
        self.shadow_hand_dof_upper_limits = to_torch(self.shadow_hand_dof_upper_limits, device=self.device)
        self.shadow_hand_dof_default_pos = to_torch(self.shadow_hand_dof_default_pos, device=self.device)
        self.shadow_hand_dof_default_vel = to_torch(self.shadow_hand_dof_default_vel, device=self.device)

        self.fingertip_handles = [self.gym.find_asset_rigid_body_index(shadow_hand_asset, name) for name in self.fingertips]

        # create fingertip force sensors, if needed
        if self.obs_type == "full_state" or self.asymmetric_obs:
            sensor_pose = gymapi.Transform()
            for ft_handle in self.fingertip_handles:
                self.gym.create_asset_force_sensor(shadow_hand_asset, ft_handle, sensor_pose)

        # load manipulated object and goal assets
        object_asset_options = gymapi.AssetOptions()
        object_asset = self.gym.load_asset(self.sim, asset_root, object_asset_file, object_asset_options)

        object_asset_options.disable_gravity = True
        goal_asset = self.gym.load_asset(self.sim, asset_root, object_asset_file, object_asset_options)

        shadow_hand_start_pose = gymapi.Transform()
        shadow_hand_start_pose.p = gymapi.Vec3(*get_axis_params(0.5, self.up_axis_idx))

        object_start_pose = gymapi.Transform()
        object_start_pose.p = gymapi.Vec3()
        object_start_pose.p.x = shadow_hand_start_pose.p.x
        pose_dy, pose_dz = -0.39, 0.10  

        object_start_pose.p.y = shadow_hand_start_pose.p.y + pose_dy
        object_start_pose.p.z = shadow_hand_start_pose.p.z + pose_dz

        if self.object_type == "pen":
            object_start_pose.p.z = shadow_hand_start_pose.p.z + 0.02

        self.goal_displacement = gymapi.Vec3(-0.2, -0.06, 0.12)
        self.goal_displacement_tensor = to_torch(
            [self.goal_displacement.x, self.goal_displacement.y, self.goal_displacement.z], device=self.device)
        goal_start_pose = gymapi.Transform()
        goal_start_pose.p = object_start_pose.p + self.goal_displacement

        goal_start_pose.p.z -= 0.04

        # compute aggregate size
        max_agg_bodies = self.num_shadow_hand_bodies + 2
        max_agg_shapes = self.num_shadow_hand_shapes + 2

        self.shadow_hands = []
        self.envs = []

        self.object_init_state = []
        self.hand_start_states = []

        self.hand_indices = []
        self.fingertip_indices = []
        self.object_indices = []
        self.goal_object_indices = []

        self.fingertip_handles = [self.gym.find_asset_rigid_body_index(shadow_hand_asset, name) for name in self.fingertips]

        shadow_hand_rb_count = self.gym.get_asset_rigid_body_count(shadow_hand_asset)
        object_rb_count = self.gym.get_asset_rigid_body_count(object_asset)
        self.object_rb_handles = list(range(shadow_hand_rb_count, shadow_hand_rb_count + object_rb_count))

        for i in range(self.num_envs):
            # create env instance
            env_ptr = self.gym.create_env(
                self.sim, lower, upper, num_per_row
            )

            if self.aggregate_mode >= 1:
                self.gym.begin_aggregate(env_ptr, max_agg_bodies, max_agg_shapes, True)

            # add hand - collision filter = -1 to use asset collision filters set in mjcf loader
            shadow_hand_actor = self.gym.create_actor(env_ptr, shadow_hand_asset, shadow_hand_start_pose, "hand", i, -1, 0)
            self.hand_start_states.append([shadow_hand_start_pose.p.x, shadow_hand_start_pose.p.y, shadow_hand_start_pose.p.z,
                                           shadow_hand_start_pose.r.x, shadow_hand_start_pose.r.y, shadow_hand_start_pose.r.z, shadow_hand_start_pose.r.w,
                                           0, 0, 0, 0, 0, 0])
            self.gym.set_actor_dof_properties(env_ptr, shadow_hand_actor, shadow_hand_dof_props)
            hand_idx = self.gym.get_actor_index(env_ptr, shadow_hand_actor, gymapi.DOMAIN_SIM)
            self.hand_indices.append(hand_idx)

            # enable DOF force sensors, if needed
            if self.obs_type == "full_state" or self.asymmetric_obs:
                self.gym.enable_actor_dof_force_sensors(env_ptr, shadow_hand_actor)

            # add object
            object_handle = self.gym.create_actor(env_ptr, object_asset, object_start_pose, "object", i, 0, 0)
            self.object_init_state.append([object_start_pose.p.x, object_start_pose.p.y, object_start_pose.p.z,
                                           object_start_pose.r.x, object_start_pose.r.y, object_start_pose.r.z, object_start_pose.r.w,
                                           0, 0, 0, 0, 0, 0])
            object_idx = self.gym.get_actor_index(env_ptr, object_handle, gymapi.DOMAIN_SIM)
            self.object_indices.append(object_idx)

            # add goal object
            goal_handle = self.gym.create_actor(env_ptr, goal_asset, goal_start_pose, "goal_object", i + self.num_envs, 0, 0)
            goal_object_idx = self.gym.get_actor_index(env_ptr, goal_handle, gymapi.DOMAIN_SIM)
            self.goal_object_indices.append(goal_object_idx)

            if self.object_type != "block":
                self.gym.set_rigid_body_color(
                    env_ptr, object_handle, 0, gymapi.MESH_VISUAL, gymapi.Vec3(0.6, 0.72, 0.98))
                self.gym.set_rigid_body_color(
                    env_ptr, goal_handle, 0, gymapi.MESH_VISUAL, gymapi.Vec3(0.6, 0.72, 0.98))

            if self.aggregate_mode > 0:
                self.gym.end_aggregate(env_ptr)

            self.envs.append(env_ptr)
            self.shadow_hands.append(shadow_hand_actor)

        # we are not using new mass values after DR when calculating random forces applied to an object,
        # which should be ok as long as the randomization range is not too big
        object_rb_props = self.gym.get_actor_rigid_body_properties(env_ptr, object_handle)
        self.object_rb_masses = [prop.mass for prop in object_rb_props]

        self.object_init_state = to_torch(self.object_init_state, device=self.device, dtype=torch.float).view(self.num_envs, 13)
        self.goal_states = self.object_init_state.clone()
        self.goal_states[:, self.up_axis_idx] -= 0.04
        self.goal_init_state = self.goal_states.clone()
        self.hand_start_states = to_torch(self.hand_start_states, device=self.device).view(self.num_envs, 13)

        self.fingertip_handles = to_torch(self.fingertip_handles, dtype=torch.long, device=self.device)
        self.object_rb_handles = to_torch(self.object_rb_handles, dtype=torch.long, device=self.device)
        self.object_rb_masses = to_torch(self.object_rb_masses, dtype=torch.float, device=self.device)

        self.hand_indices = to_torch(self.hand_indices, dtype=torch.long, device=self.device)
        self.object_indices = to_torch(self.object_indices, dtype=torch.long, device=self.device)
        self.goal_object_indices = to_torch(self.goal_object_indices, dtype=torch.long, device=self.device)

    def compute_reward(self, actions):
        self.rew_buf[:], self.reset_buf[:], self.reset_goal_buf[:], self.progress_buf[:], self.successes[:], self.consecutive_successes[:] = compute_hand_reward(
            self.rew_buf, self.reset_buf, self.reset_goal_buf, self.progress_buf, self.successes, self.consecutive_successes,
            self.max_episode_length, self.object_pos, self.object_rot, self.goal_pos, self.goal_rot,
            self.dist_reward_scale, self.rot_reward_scale, self.rot_eps, self.actions, self.action_penalty_scale,
            self.success_tolerance, self.reach_goal_bonus, self.fall_dist, self.fall_penalty,
            self.max_consecutive_successes, self.av_factor, (self.object_type == "pen")
        )

        self.extras['consecutive_successes'] = self.consecutive_successes.mean()

        if self.print_success_stat:
            self.total_resets = self.total_resets + self.reset_buf.sum()
            direct_average_successes = self.total_successes + self.successes.sum()
            self.total_successes = self.total_successes + (self.successes * self.reset_buf).sum()

            # The direct average shows the overall result more quickly, but slightly undershoots long term
            # policy performance.
            print("Direct average consecutive successes = {:.1f}".format(direct_average_successes/(self.total_resets + self.num_envs)))
            if self.total_resets > 0:
                print("Post-Reset average consecutive successes = {:.1f}".format(self.total_successes/self.total_resets))

    def compute_observations(self):
        self.gym.refresh_dof_state_tensor(self.sim)
        self.gym.refresh_actor_root_state_tensor(self.sim)
        self.gym.refresh_rigid_body_state_tensor(self.sim)

        if self.obs_type == "full_state" or self.asymmetric_obs:
            self.gym.refresh_force_sensor_tensor(self.sim)
            self.gym.refresh_dof_force_tensor(self.sim)

        self.object_pose = self.root_state_tensor[self.object_indices, 0:7]
        self.object_pos = self.root_state_tensor[self.object_indices, 0:3]
        self.object_rot = self.root_state_tensor[self.object_indices, 3:7]
        self.object_linvel = self.root_state_tensor[self.object_indices, 7:10]
        self.object_angvel = self.root_state_tensor[self.object_indices, 10:13]

        self.goal_pose = self.goal_states[:, 0:7]
        self.goal_pos = self.goal_states[:, 0:3]
        self.goal_rot = self.goal_states[:, 3:7]

        self.fingertip_state = self.rigid_body_states[:, self.fingertip_handles][:, :, 0:13]
        self.fingertip_pos = self.rigid_body_states[:, self.fingertip_handles][:, :, 0:3]

        if self.obs_type == "openai":
            self.compute_fingertip_observations(True)
        elif self.obs_type == "full_no_vel":
            self.compute_full_observations(True)
        elif self.obs_type == "full":
            self.compute_full_observations()
        elif self.obs_type == "full_state":
            self.compute_full_state()
        else:
            print("Unknown observations type!")

        if self.asymmetric_obs:
            self.compute_full_state(True)

    def compute_fingertip_observations(self, no_vel=False):
        if no_vel:
            # Per https://arxiv.org/pdf/1808.00177.pdf Table 2
            #   Fingertip positions
            #   Object Position, but not orientation
            #   Relative target orientation

            # 3*self.num_fingertips = 15
            self.obs_buf[:, 0:15] = self.fingertip_pos.reshape(self.num_envs, 15)
            self.obs_buf[:, 15:18] = self.object_pose[:, 0:3]
            self.obs_buf[:, 18:22] = quat_mul(self.object_rot, quat_conjugate(self.goal_rot))

            self.obs_buf[:, 22:42] = self.actions
        else:
            # 13*self.num_fingertips = 65
            self.obs_buf[:, 0:65] = self.fingertip_state.reshape(self.num_envs, 65)
            self.obs_buf[:, 65:72] = self.object_pose
            self.obs_buf[:, 72:75] = self.object_linvel
            self.obs_buf[:, 75:78] = self.vel_obs_scale * self.object_angvel

            self.obs_buf[:, 78:85] = self.goal_pose
            self.obs_buf[:, 85:89] = quat_mul(self.object_rot, quat_conjugate(self.goal_rot))

            self.obs_buf[:, 89:109] = self.actions

    def compute_full_observations(self, no_vel=False):
        if no_vel:
            self.obs_buf[:, 0:self.num_shadow_hand_dofs] = unscale(self.shadow_hand_dof_pos,
                                                                   self.shadow_hand_dof_lower_limits, self.shadow_hand_dof_upper_limits)

            self.obs_buf[:, 24:31] = self.object_pose
            self.obs_buf[:, 31:38] = self.goal_pose
            self.obs_buf[:, 38:42] = quat_mul(self.object_rot, quat_conjugate(self.goal_rot))

            # 3*self.num_fingertips = 15
            self.obs_buf[:, 42:57] = self.fingertip_pos.reshape(self.num_envs, 15)

            self.obs_buf[:, 57:77] = self.actions
        else:
            self.obs_buf[:, 0:self.num_shadow_hand_dofs] = unscale(self.shadow_hand_dof_pos,
                                                                   self.shadow_hand_dof_lower_limits, self.shadow_hand_dof_upper_limits)
            self.obs_buf[:, self.num_shadow_hand_dofs:2*self.num_shadow_hand_dofs] = self.vel_obs_scale * self.shadow_hand_dof_vel

            self.obs_buf[:, 48:55] = self.object_pose
            self.obs_buf[:, 55:58] = self.object_linvel
            self.obs_buf[:, 58:61] = self.vel_obs_scale * self.object_angvel

            self.obs_buf[:, 61:68] = self.goal_pose
            self.obs_buf[:, 68:72] = quat_mul(self.object_rot, quat_conjugate(self.goal_rot))

            # 13*self.num_fingertips = 65
            self.obs_buf[:, 72:137] = self.fingertip_state.reshape(self.num_envs, 65)

            self.obs_buf[:, 137:157] = self.actions

    def compute_full_state(self, asymm_obs=False):
        if asymm_obs:
            self.states_buf[:, 0:self.num_shadow_hand_dofs] = unscale(self.shadow_hand_dof_pos,
                                                                      self.shadow_hand_dof_lower_limits, self.shadow_hand_dof_upper_limits)
            self.states_buf[:, self.num_shadow_hand_dofs:2*self.num_shadow_hand_dofs] = self.vel_obs_scale * self.shadow_hand_dof_vel
            self.states_buf[:, 2*self.num_shadow_hand_dofs:3*self.num_shadow_hand_dofs] = self.force_torque_obs_scale * self.dof_force_tensor

            obj_obs_start = 3*self.num_shadow_hand_dofs  # 72
            self.states_buf[:, obj_obs_start:obj_obs_start + 7] = self.object_pose
            self.states_buf[:, obj_obs_start + 7:obj_obs_start + 10] = self.object_linvel
            self.states_buf[:, obj_obs_start + 10:obj_obs_start + 13] = self.vel_obs_scale * self.object_angvel

            goal_obs_start = obj_obs_start + 13  # 85
            self.states_buf[:, goal_obs_start:goal_obs_start + 7] = self.goal_pose
            self.states_buf[:, goal_obs_start + 7:goal_obs_start + 11] = quat_mul(self.object_rot, quat_conjugate(self.goal_rot))

            # fingertip observations, state(pose and vel) + force-torque sensors
            num_ft_states = 13 * self.num_fingertips  # 65
            num_ft_force_torques = 6 * self.num_fingertips  # 30

            fingertip_obs_start = goal_obs_start + 11  # 96
            self.states_buf[:, fingertip_obs_start:fingertip_obs_start + num_ft_states] = self.fingertip_state.reshape(self.num_envs, num_ft_states)
            self.states_buf[:, fingertip_obs_start + num_ft_states:fingertip_obs_start + num_ft_states +
                            num_ft_force_torques] = self.force_torque_obs_scale * self.vec_sensor_tensor

            # obs_end = 96 + 65 + 30 = 191
            # obs_total = obs_end + num_actions = 211
            obs_end = fingertip_obs_start + num_ft_states + num_ft_force_torques
            self.states_buf[:, obs_end:obs_end + self.num_actions] = self.actions
        else:
            self.obs_buf[:, 0:self.num_shadow_hand_dofs] = unscale(self.shadow_hand_dof_pos,
                                                                   self.shadow_hand_dof_lower_limits, self.shadow_hand_dof_upper_limits)
            self.obs_buf[:, self.num_shadow_hand_dofs:2*self.num_shadow_hand_dofs] = self.vel_obs_scale * self.shadow_hand_dof_vel
            self.obs_buf[:, 2*self.num_shadow_hand_dofs:3*self.num_shadow_hand_dofs] = self.force_torque_obs_scale * self.dof_force_tensor

            obj_obs_start = 3*self.num_shadow_hand_dofs  # 72
            self.obs_buf[:, obj_obs_start:obj_obs_start + 7] = self.object_pose
            self.obs_buf[:, obj_obs_start + 7:obj_obs_start + 10] = self.object_linvel
            self.obs_buf[:, obj_obs_start + 10:obj_obs_start + 13] = self.vel_obs_scale * self.object_angvel

            goal_obs_start = obj_obs_start + 13  # 85
            self.obs_buf[:, goal_obs_start:goal_obs_start + 7] = self.goal_pose
            self.obs_buf[:, goal_obs_start + 7:goal_obs_start + 11] = quat_mul(self.object_rot, quat_conjugate(self.goal_rot))

            # fingertip observations, state(pose and vel) + force-torque sensors
            num_ft_states = 13 * self.num_fingertips  # 65
            num_ft_force_torques = 6 * self.num_fingertips  # 30

            fingertip_obs_start = goal_obs_start + 11  # 96
            self.obs_buf[:, fingertip_obs_start:fingertip_obs_start + num_ft_states] = self.fingertip_state.reshape(self.num_envs, num_ft_states)
            self.obs_buf[:, fingertip_obs_start + num_ft_states:fingertip_obs_start + num_ft_states +
                         num_ft_force_torques] = self.force_torque_obs_scale * self.vec_sensor_tensor

            # obs_end = 96 + 65 + 30 = 191
            # obs_total = obs_end + num_actions = 211
            obs_end = fingertip_obs_start + num_ft_states + num_ft_force_torques
            self.obs_buf[:, obs_end:obs_end + self.num_actions] = self.actions

    def reset_target_pose(self, env_ids, apply_reset=False):
        rand_floats = torch_rand_float(-1.0, 1.0, (len(env_ids), 4), device=self.device)

        new_rot = randomize_rotation(rand_floats[:, 0], rand_floats[:, 1], self.x_unit_tensor[env_ids], self.y_unit_tensor[env_ids])

        self.goal_states[env_ids, 0:3] = self.goal_init_state[env_ids, 0:3]
        self.goal_states[env_ids, 3:7] = new_rot
        self.root_state_tensor[self.goal_object_indices[env_ids], 0:3] = self.goal_states[env_ids, 0:3] + self.goal_displacement_tensor
        self.root_state_tensor[self.goal_object_indices[env_ids], 3:7] = self.goal_states[env_ids, 3:7]
        self.root_state_tensor[self.goal_object_indices[env_ids], 7:13] = torch.zeros_like(self.root_state_tensor[self.goal_object_indices[env_ids], 7:13])

        if apply_reset:
            goal_object_indices = self.goal_object_indices[env_ids].to(torch.int32)
            self.gym.set_actor_root_state_tensor_indexed(self.sim,
                                                         gymtorch.unwrap_tensor(self.root_state_tensor),
                                                         gymtorch.unwrap_tensor(goal_object_indices), len(env_ids))
        self.reset_goal_buf[env_ids] = 0

    def reset_idx(self, env_ids, goal_env_ids):
        # randomization can happen only at reset time, since it can reset actor positions on GPU
        if self.randomize:
            self.apply_randomizations(self.randomization_params)

        # generate random values
        rand_floats = torch_rand_float(-1.0, 1.0, (len(env_ids), self.num_shadow_hand_dofs * 2 + 5), device=self.device)

        # randomize start object poses
        self.reset_target_pose(env_ids)

        # reset rigid body forces
        self.rb_forces[env_ids, :, :] = 0.0

        # reset object
        self.root_state_tensor[self.object_indices[env_ids]] = self.object_init_state[env_ids].clone()
        self.root_state_tensor[self.object_indices[env_ids], 0:2] = self.object_init_state[env_ids, 0:2] + \
            self.reset_position_noise * rand_floats[:, 0:2]
        self.root_state_tensor[self.object_indices[env_ids], self.up_axis_idx] = self.object_init_state[env_ids, self.up_axis_idx] + \
            self.reset_position_noise * rand_floats[:, self.up_axis_idx]

        new_object_rot = randomize_rotation(rand_floats[:, 3], rand_floats[:, 4], self.x_unit_tensor[env_ids], self.y_unit_tensor[env_ids])
        if self.object_type == "pen":
            rand_angle_y = torch.tensor(0.3)
            new_object_rot = randomize_rotation_pen(rand_floats[:, 3], rand_floats[:, 4], rand_angle_y,
                                                    self.x_unit_tensor[env_ids], self.y_unit_tensor[env_ids], self.z_unit_tensor[env_ids])

        self.root_state_tensor[self.object_indices[env_ids], 3:7] = new_object_rot
        self.root_state_tensor[self.object_indices[env_ids], 7:13] = torch.zeros_like(self.root_state_tensor[self.object_indices[env_ids], 7:13])

        object_indices = torch.unique(torch.cat([self.object_indices[env_ids],
                                                 self.goal_object_indices[env_ids],
                                                 self.goal_object_indices[goal_env_ids]]).to(torch.int32))
        self.gym.set_actor_root_state_tensor_indexed(self.sim,
                                                     gymtorch.unwrap_tensor(self.root_state_tensor),
                                                     gymtorch.unwrap_tensor(object_indices), len(object_indices))

        # reset random force probabilities
        self.random_force_prob[env_ids] = torch.exp((torch.log(self.force_prob_range[0]) - torch.log(self.force_prob_range[1]))
                                                    * torch.rand(len(env_ids), device=self.device) + torch.log(self.force_prob_range[1]))

        # reset shadow hand
        delta_max = self.shadow_hand_dof_upper_limits - self.shadow_hand_dof_default_pos
        delta_min = self.shadow_hand_dof_lower_limits - self.shadow_hand_dof_default_pos
        rand_delta = delta_min + (delta_max - delta_min) * 0.5 * (rand_floats[:, 5:5+self.num_shadow_hand_dofs] + 1)

        pos = self.shadow_hand_default_dof_pos + self.reset_dof_pos_noise * rand_delta
        self.shadow_hand_dof_pos[env_ids, :] = pos
        self.shadow_hand_dof_vel[env_ids, :] = self.shadow_hand_dof_default_vel + \
            self.reset_dof_vel_noise * rand_floats[:, 5+self.num_shadow_hand_dofs:5+self.num_shadow_hand_dofs*2]
        self.prev_targets[env_ids, :self.num_shadow_hand_dofs] = pos
        self.cur_targets[env_ids, :self.num_shadow_hand_dofs] = pos

        hand_indices = self.hand_indices[env_ids].to(torch.int32)
        self.gym.set_dof_position_target_tensor_indexed(self.sim,
                                                        gymtorch.unwrap_tensor(self.prev_targets),
                                                        gymtorch.unwrap_tensor(hand_indices), len(env_ids))

        self.gym.set_dof_state_tensor_indexed(self.sim,
                                              gymtorch.unwrap_tensor(self.dof_state),
                                              gymtorch.unwrap_tensor(hand_indices), len(env_ids))

        self.progress_buf[env_ids] = 0
        self.reset_buf[env_ids] = 0
        self.successes[env_ids] = 0

    def pre_physics_step(self, actions):
        env_ids = self.reset_buf.nonzero(as_tuple=False).squeeze(-1)
        goal_env_ids = self.reset_goal_buf.nonzero(as_tuple=False).squeeze(-1)

        # if only goals need reset, then call set API
        if len(goal_env_ids) > 0 and len(env_ids) == 0:
            self.reset_target_pose(goal_env_ids, apply_reset=True)
        # if goals need reset in addition to other envs, call set API in reset_idx()
        elif len(goal_env_ids) > 0:
            self.reset_target_pose(goal_env_ids)

        if len(env_ids) > 0:
            self.reset_idx(env_ids, goal_env_ids)

        self.actions = actions.clone().to(self.device)
        if self.use_relative_control:
            targets = self.prev_targets[:, self.actuated_dof_indices] + self.shadow_hand_dof_speed_scale * self.dt * self.actions
            self.cur_targets[:, self.actuated_dof_indices] = tensor_clamp(targets,
                                                                          self.shadow_hand_dof_lower_limits[self.actuated_dof_indices], self.shadow_hand_dof_upper_limits[self.actuated_dof_indices])
        else:
            self.cur_targets[:, self.actuated_dof_indices] = scale(self.actions,
                                                                   self.shadow_hand_dof_lower_limits[self.actuated_dof_indices], self.shadow_hand_dof_upper_limits[self.actuated_dof_indices])
            self.cur_targets[:, self.actuated_dof_indices] = self.act_moving_average * self.cur_targets[:,
                                                                                                        self.actuated_dof_indices] + (1.0 - self.act_moving_average) * self.prev_targets[:, self.actuated_dof_indices]
            self.cur_targets[:, self.actuated_dof_indices] = tensor_clamp(self.cur_targets[:, self.actuated_dof_indices],
                                                                          self.shadow_hand_dof_lower_limits[self.actuated_dof_indices], self.shadow_hand_dof_upper_limits[self.actuated_dof_indices])

        self.prev_targets[:, self.actuated_dof_indices] = self.cur_targets[:, self.actuated_dof_indices]
        self.gym.set_dof_position_target_tensor(self.sim, gymtorch.unwrap_tensor(self.cur_targets))

        if self.force_scale > 0.0:
            self.rb_forces *= torch.pow(self.force_decay, self.dt / self.force_decay_interval)

            # apply new forces
            force_indices = (torch.rand(self.num_envs, device=self.device) < self.random_force_prob).nonzero()
            self.rb_forces[force_indices, self.object_rb_handles, :] = torch.randn(
                self.rb_forces[force_indices, self.object_rb_handles, :].shape, device=self.device) * self.object_rb_masses * self.force_scale

            self.gym.apply_rigid_body_force_tensors(self.sim, gymtorch.unwrap_tensor(self.rb_forces), None, gymapi.LOCAL_SPACE)

    def post_physics_step(self):
        self.progress_buf += 1
        self.randomize_buf += 1

        self.compute_observations()
        self.compute_reward(self.actions)

        if self.viewer and self.debug_viz:
            # draw axes on target object
            self.gym.clear_lines(self.viewer)
            self.gym.refresh_rigid_body_state_tensor(self.sim)

            for i in range(self.num_envs):
                targetx = (self.goal_pos[i] + quat_apply(self.goal_rot[i], to_torch([1, 0, 0], device=self.device) * 0.2)).cpu().numpy()
                targety = (self.goal_pos[i] + quat_apply(self.goal_rot[i], to_torch([0, 1, 0], device=self.device) * 0.2)).cpu().numpy()
                targetz = (self.goal_pos[i] + quat_apply(self.goal_rot[i], to_torch([0, 0, 1], device=self.device) * 0.2)).cpu().numpy()

                p0 = self.goal_pos[i].cpu().numpy() + self.goal_displacement_tensor.cpu().numpy()
                self.gym.add_lines(self.viewer, self.envs[i], 1, [p0[0], p0[1], p0[2], targetx[0], targetx[1], targetx[2]], [0.85, 0.1, 0.1])
                self.gym.add_lines(self.viewer, self.envs[i], 1, [p0[0], p0[1], p0[2], targety[0], targety[1], targety[2]], [0.1, 0.85, 0.1])
                self.gym.add_lines(self.viewer, self.envs[i], 1, [p0[0], p0[1], p0[2], targetz[0], targetz[1], targetz[2]], [0.1, 0.1, 0.85])

                objectx = (self.object_pos[i] + quat_apply(self.object_rot[i], to_torch([1, 0, 0], device=self.device) * 0.2)).cpu().numpy()
                objecty = (self.object_pos[i] + quat_apply(self.object_rot[i], to_torch([0, 1, 0], device=self.device) * 0.2)).cpu().numpy()
                objectz = (self.object_pos[i] + quat_apply(self.object_rot[i], to_torch([0, 0, 1], device=self.device) * 0.2)).cpu().numpy()

                p0 = self.object_pos[i].cpu().numpy()
                self.gym.add_lines(self.viewer, self.envs[i], 1, [p0[0], p0[1], p0[2], objectx[0], objectx[1], objectx[2]], [0.85, 0.1, 0.1])
                self.gym.add_lines(self.viewer, self.envs[i], 1, [p0[0], p0[1], p0[2], objecty[0], objecty[1], objecty[2]], [0.1, 0.85, 0.1])
                self.gym.add_lines(self.viewer, self.envs[i], 1, [p0[0], p0[1], p0[2], objectz[0], objectz[1], objectz[2]], [0.1, 0.1, 0.85])

#####################################################################
###=========================jit functions=========================###
#####################################################################


@torch.jit.script
def compute_hand_reward(
    rew_buf, reset_buf, reset_goal_buf, progress_buf, successes, consecutive_successes,
    max_episode_length: float, object_pos, object_rot, target_pos, target_rot,
    dist_reward_scale: float, rot_reward_scale: float, rot_eps: float,
    actions, action_penalty_scale: float,
    success_tolerance: float, reach_goal_bonus: float, fall_dist: float,
    fall_penalty: float, max_consecutive_successes: int, av_factor: float, ignore_z_rot: bool
):
    # Distance from the hand to the object
    goal_dist = torch.norm(object_pos - target_pos, p=2, dim=-1)

    if ignore_z_rot:
        success_tolerance = 2.0 * success_tolerance

    # Orientation alignment for the cube in hand and goal cube
    quat_diff = quat_mul(object_rot, quat_conjugate(target_rot))
    rot_dist = 2.0 * torch.asin(torch.clamp(torch.norm(quat_diff[:, 0:3], p=2, dim=-1), max=1.0))

    dist_rew = goal_dist * dist_reward_scale
    rot_rew = 1.0/(torch.abs(rot_dist) + rot_eps) * rot_reward_scale

    action_penalty = torch.sum(actions ** 2, dim=-1)

    # Total reward is: position distance + orientation alignment + action regularization + success bonus + fall penalty
    reward = dist_rew + rot_rew + action_penalty * action_penalty_scale

    # Find out which envs hit the goal and update successes count
    goal_resets = torch.where(torch.abs(rot_dist) <= success_tolerance, torch.ones_like(reset_goal_buf), reset_goal_buf)
    successes = successes + goal_resets

    # Success bonus: orientation is within `success_tolerance` of goal orientation
    reward = torch.where(goal_resets == 1, reward + reach_goal_bonus, reward)

    # Fall penalty: distance to the goal is larger than a threshold
    reward = torch.where(goal_dist >= fall_dist, reward + fall_penalty, reward)

    # Check env termination conditions, including maximum success number
    resets = torch.where(goal_dist >= fall_dist, torch.ones_like(reset_buf), reset_buf)
    if max_consecutive_successes > 0:
        # Reset progress buffer on goal envs if max_consecutive_successes > 0
        progress_buf = torch.where(torch.abs(rot_dist) <= success_tolerance, torch.zeros_like(progress_buf), progress_buf)
        resets = torch.where(successes >= max_consecutive_successes, torch.ones_like(resets), resets)
    resets = torch.where(progress_buf >= max_episode_length - 1, torch.ones_like(resets), resets)

    # Apply penalty for not reaching the goal
    if max_consecutive_successes > 0:
        reward = torch.where(progress_buf >= max_episode_length - 1, reward + 0.5 * fall_penalty, reward)

    num_resets = torch.sum(resets)
    finished_cons_successes = torch.sum(successes * resets.float())

    cons_successes = torch.where(num_resets > 0, av_factor*finished_cons_successes/num_resets + (1.0 - av_factor)*consecutive_successes, consecutive_successes)

    return reward, resets, goal_resets, progress_buf, successes, cons_successes


@torch.jit.script
def randomize_rotation(rand0, rand1, x_unit_tensor, y_unit_tensor):
    return quat_mul(quat_from_angle_axis(rand0 * np.pi, x_unit_tensor),
                    quat_from_angle_axis(rand1 * np.pi, y_unit_tensor))


@torch.jit.script
def randomize_rotation_pen(rand0, rand1, max_angle, x_unit_tensor, y_unit_tensor, z_unit_tensor):
    rot = quat_mul(quat_from_angle_axis(0.5 * np.pi + rand0 * max_angle, x_unit_tensor),
                   quat_from_angle_axis(rand0 * np.pi, z_unit_tensor))
    return rot