""" @author: Olivier Sigaud A merge between two sources: * Adaptation of the MountainCar Environment from the "FAReinforcement" library of Jose Antonio Martin H. (version 1.0), adapted by 'Tom Schaul, tom@idsia.ch' and then modified by Arnaud de Broissia * the gym MountainCar environment itself from http://incompleteideas.net/sutton/MountainCar/MountainCar1.cp permalink: https://perma.cc/6Z2N-PFWC """ # apple: https://unsplash.com/images/food/apple # orange: https://unsplash.com/s/photos/orange # wood: https://architextures.org/textures/category/wood import math from typing import Optional import numpy as np import gym from gym import spaces # from gym.envs.classic_control import utils from gym.error import DependencyNotInstalled # from gym.utils.renderer import Renderer import pygame import scipy import yaml from collections import OrderedDict import copy class TinyUR5Env(gym.Env): """ ### Description The Mountain Car MDP is a deterministic MDP that consists of a car placed stochastically at the bottom of a sinusoidal valley, with the only possible actions being the accelerations that can be applied to the car in either direction. The goal of the MDP is to strategically accelerate the car to reach the goal state on top of the right hill. There are two versions of the mountain car domain in gym: one with discrete actions and one with continuous. This version is the one with continuous actions. This MDP first appeared in [Andrew Moore's PhD Thesis (1990)](https://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-209.pdf) ``` @TECHREPORT{Moore90efficientmemory-based, author = {Andrew William Moore}, title = {Efficient Memory-based Learning for Robot Control}, institution = {University of Cambridge}, year = {1990} } ``` ### Observation Space The observation is a `ndarray` with shape `(2,)` where the elements correspond to the following: | Num | Observation | Min | Max | Unit | |-----|--------------------------------------|------|-----|--------------| | 0 | position of the car along the x-axis | -Inf | Inf | position (m) | | 1 | velocity of the car | -Inf | Inf | position (m) | ### Action Space The action is a `ndarray` with shape `(1,)`, representing the directional force applied on the car. The action is clipped in the range `[-1,1]` and multiplied by a power of 0.0015. ### Transition Dynamics: Given an action, the mountain car follows the following transition dynamics: *velocity_t+1 = velocity_t+1 + force * self.power - 0.0025 * cos(3 * position_t)* *position_t+1 = position_t + velocity_t+1* where force is the action clipped to the range `[-1,1]` and power is a constant 0.0015. The collisions at either end are inelastic with the velocity set to 0 upon collision with the wall. The position is clipped to the range [-1.2, 0.6] and velocity is clipped to the range [-0.07, 0.07]. ### Reward A negative reward of *-0.1 * action²* is received at each timestep to penalise for taking actions of large magnitude. If the mountain car reaches the goal then a positive reward of +100 is added to the negative reward for that timestep. ### Starting State The position of the car is assigned a uniform random value in `[-0.6 , -0.4]`. The starting velocity of the car is always assigned to 0. ### Episode End The episode ends if either of the following happens: 1. Termination: The position of the car is greater than or equal to 0.45 (the goal position on top of the right hill) 2. Truncation: The length of the episode is 999. ### Arguments ``` gym.make('MountainCarContinuous-v0') ``` ### Version History * v0: Initial versions release (1.0.0) """ metadata = { "render_modes": ["human", "rgb_array", "single_rgb_array"], "render_fps": 30, } # def __init__(self, yaml_file='config.yaml', render_mode: Optional[str] = None, goal_velocity=0, initializer=None): def __init__(self, config, render_mode: Optional[str] = None, goal_velocity=0): # with open(yaml_file, "r") as stream: # try: # config = yaml.safe_load(stream) # # print(config, type(config)) # except yaml.YAMLError as exc: # print(exc) # # exit() # if initializer is not None: # config = initializer.initialize self.config = config self.min_action = -np.pi * 2 - 0.01 self.max_action = np.pi * 2 + 0.01 self.min_position = -1.2 self.max_position = 0.6 self.max_speed = 0.07 self.goal_position = ( 0.45 # was 0.5 in gym, 0.45 in Arnaud de Broissia's version ) self.goal_velocity = goal_velocity self.power = 0.0015 self.low_state = np.array( [self.min_position, -self.max_speed], dtype=np.float32 ) self.high_state = np.array( [self.max_position, self.max_speed], dtype=np.float32 ) self.render_mode = render_mode # self.renderer = Renderer(self.render_mode, self._render) self.scale = config['scale'] self.screen_width = int(config['desk_width'] * self.scale) self.screen_height = int(config['desk_height'] * self.scale) self.robot_base_xy = [config['robot']['base_x'] * self.scale, config['robot']['base_y'] * self.scale] self.tool_center_point = config['robot']['tool_center_point_distance'] * self.scale self.tool_img_mid_point = config['robot']['tool_img_mid_point'] * self.scale self.screen = None self.clock = None self.isopen = True self.action_space = spaces.Box( low=self.min_action, high=self.max_action, shape=(4,), dtype=np.float32 ) self.observation_space = spaces.Box( low=self.min_action, high=self.max_action, shape=(4,), dtype=np.float32 ) if 'init_joints' not in config: self.robot_joints = np.zeros((4,), dtype=np.float32) self.robot_joints[0] = -1.57 self.robot_joints[1] = 1.57 self.robot_joints[2] = 0 self.robot_joints_init = np.zeros((4,), dtype=np.float32) self.robot_joints_init[0] = -1.57 self.robot_joints_init[1] = 1.57 self.robot_joints_init[2] = 0 else: self.robot_joints = np.zeros((4,), dtype=np.float32) for i in range(4): self.robot_joints[i] = config['init_joints'][i] self.robot_joints_init = copy.deepcopy(self.robot_joints) self.lim_length = config['objects']['lim']['length'] * self.scale self.Kp = 15 self.dt = 0.003 self.manip_objs = OrderedDict() self.env_objs = {} for obj in config['objects']: # print(obj) # print(config['objects'][obj]) obj_img = pygame.image.load(config['objects'][obj]['image']) if 'position' not in config['objects'][obj]: self.env_objs[obj] = {} self.env_objs[obj]['size_xy'] = [config['objects'][obj]['size']['x'] * self.scale, config['objects'][obj]['size']['y'] * self.scale] self.env_objs[obj]['image'] = \ pygame.transform.smoothscale( obj_img, self.env_objs[obj]['size_xy']) if 'position' in config['objects'][obj]: self.manip_objs[obj] = {} self.manip_objs[obj]['size_xy'] = [config['objects'][obj]['size']['x'] * self.scale, config['objects'][obj]['size']['y'] * self.scale] self.manip_objs[obj]['pos_xy'] = [config['objects'][obj]['position']['x'] * self.scale, config['objects'][obj]['position']['y'] * self.scale] self.manip_objs[obj]['pos_z'] = config['objects'][obj]['position']['z'] * self.scale self.manip_objs[obj]['size_z'] = config['objects'][obj]['size']['z'] * self.scale self.manip_objs[obj]['orientation'] = 0. self.manip_objs[obj]['lru_score'] = 0 self.manip_objs[obj]['image'] = \ pygame.transform.smoothscale( obj_img, self.manip_objs[obj]['size_xy']) # exit() # print(self.manip_objs) self.eef_z = 120 * self.scale self.grab = None self.grab_position = None self.grab_position_z = None self.grab_orientation = None self.highest_lru_score = 0 def _eef_(self): start_x = 0 start_y = 0 end_x = self.robot_base_xy[0] end_y = self.robot_base_xy[1] angle = 0 for i in range(self.robot_joints.shape[0] - 1): if i < 2: start_x = end_x start_y = end_y angle = angle + self.robot_joints[i] end_x = start_x + np.sin(angle) * self.lim_length end_y = start_y + np.cos(angle) * self.lim_length # mid_x = (start_x + end_x) / 2 # mid_y = (start_y + end_y) / 2 elif i == 2: start_x = end_x start_y = end_y angle = angle + self.robot_joints[i] end_x = start_x + np.sin(angle) * self.tool_center_point end_y = start_y + np.cos(angle) * self.tool_center_point # mid_x = (start_x + end_x) / 2 # mid_y = (start_y + end_y) / 2 return np.array([end_x, end_y]) def _eef_orientation_(self): # print('joints0', self.robot_joints[0], self.robot_joints[1], self.robot_joints[2]) # print('joints1', float(self.robot_joints[0]), float(self.robot_joints[1]), float(self.robot_joints[2])) return float(self.robot_joints[0]) + float(self.robot_joints[1]) + float(self.robot_joints[2]) # return float(self.robot_joints[0] + self.robot_joints[1] + self.robot_joints[2]) def _l2_(self, eef, position): return ((eef[0] - position[0]) ** 2 + (eef[1] - position[1]) ** 2) ** (1/2) def _grab_(self, position, eef): grab = (self._l2_(eef, position) < 50 * self.scale) # print(self._l2_(eef, position)) return grab def _grab_z_(self, env_obj, eef_z): lower_bound = env_obj['pos_z'] upper_bound = env_obj['pos_z'] + env_obj['size_z'] # print(f'bounds: {lower_bound}, {upper_bound}') if eef_z >= lower_bound and eef_z <= upper_bound: return True else: return False def _gripper_closed_(self): if self.robot_joints[-1] >= 0: return True else: return False def _get_serialized_objs_(self): objs = OrderedDict() keys = [ 'size_xy', 'pos_xy', 'pos_z', 'size_z', 'orientation', 'lru_score', ] for obj in self.manip_objs: objs[obj] = {} for key in keys: objs[obj][key] = self.manip_objs[obj][key] return objs def _max_speed_(self, disp, lim): if disp > lim: return lim if disp < -lim: return -lim return disp def step(self, action: np.ndarray, eef_z=None): # print([self.manip_objs[env_obj]['lru_score'] for env_obj in self.manip_objs if 'pos_xy' in self.manip_objs[env_obj]]) # print(action, self.robot_joints) # Convert a possible numpy bool to a Python bool. terminated = False reward = 0 # action is target angles of the joints assert action.shape[0] == self.robot_joints.shape[0] for i in range(action.shape[0]): # print('before change', self.robot_joints[i]) # self.robot_joints[i] = self.robot_joints[i] + self.Kp * self._ang_diff(action[i], self.robot_joints[i]) * self.dt # self.robot_joints[i] = self.robot_joints[i] + self._max_speed_( # self._ang_diff(action[i], self.robot_joints[i]), self.Kp * self.dt) # print('after change', self.robot_joints[i]) self.robot_joints[i] = self.robot_joints[i] + self.Kp * 2 * self._ang_diff(action[i], self.robot_joints[i]) * self.dt eef = self._eef_() if eef_z is not None: self.eef_z = eef_z * self.scale # print(f'eef_z: {self.eef_z}') # print('grab', self.grab) if self.grab is not None: assert self.grab_orientation is not None self.manip_objs[self.grab]['pos_xy'] = eef + self.grab_position self.manip_objs[self.grab]['pos_z'] = self.eef_z + self.grab_position_z # print('grab orientation', self.grab_orientation) # print('eef orientation', self._eef_orientation_()) self.manip_objs[self.grab]['orientation'] = self._eef_orientation_() + self.grab_orientation # print('orientation', self.manip_objs[self.grab]['orientation']) if self.manip_objs[self.grab]['pos_z'] < 0: self.manip_objs[self.grab]['pos_z'] = 0 self.grab_position_z = self.manip_objs[self.grab]['pos_z'] - self.eef_z # print(self.grab_position, eef, self.positions[1], 1) # self.grab = None # self.grab_position = None if self._gripper_closed_(): if self.grab is None: for obj in self.manip_objs: if 'pos_xy' not in self.manip_objs[obj]: continue if self._grab_(self.manip_objs[obj]['pos_xy'], eef) and self._grab_z_(self.manip_objs[obj], self.eef_z): self.grab = obj self.grab_position = self.manip_objs[obj]['pos_xy'] - eef self.grab_position_z = self.manip_objs[self.grab]['pos_z'] - self.eef_z assert self.manip_objs[self.grab]['orientation'] is not None self.grab_orientation = self.manip_objs[self.grab]['orientation'] - self._eef_orientation_() self.highest_lru_score += 1 self.manip_objs[obj]['lru_score'] = self.highest_lru_score self.manip_objs.move_to_end(obj, last=False) break else: self.grab = None self.grab_position = None self.grab_position_z = None self.grab_orientation = None # print(self.grab_position, eef, self.positions[1], 2) # self.renderer.render_step() state = { 'joints': copy.deepcopy(self.robot_joints), 'eef': copy.deepcopy(self._eef_()), 'eef_z': copy.deepcopy(self.eef_z), 'eef_orientation': copy.deepcopy(self._eef_orientation_()), 'positions': copy.deepcopy(self._get_serialized_objs_()), 'grabbed_object': copy.deepcopy(self.grab), 'grab_position': copy.deepcopy(self.grab_position), 'grab_position_z': copy.deepcopy(self.grab_position_z), 'grab_orientation': copy.deepcopy(self.grab_orientation), 'scale': copy.deepcopy(self.scale), } return state, reward, terminated, {} def reset( self, *, seed: Optional[int] = None, return_info: bool = False, options: Optional[dict] = None ): super().reset(seed=seed) # Note that if you use custom reset bounds, it may lead to out-of-bound # state/observations. # low, high = utils.maybe_parse_reset_bounds(options, -0.6, -0.4) # self.state = np.array([self.np_random.uniform(low=low, high=high), 0]) self.state = np.array([self.np_random.uniform(low=-0.6, high=-0.4), 0]) # self.renderer.reset() # self.renderer.render_step() if not return_info: return np.array(self.state, dtype=np.float32) else: return np.array(self.state, dtype=np.float32), {} def _height(self, xs): return np.sin(3 * xs) * 0.45 + 0.55 def _ang_diff(self, theta1, theta2): # Returns the difference between two angles in the range -pi to +pi # print('angle diff', (theta1 - theta2 + np.pi) % (2 * np.pi) - np.pi) # print(f'theta1 {theta1} theta2 {theta2}') return (theta1 - theta2 + np.pi) % (2 * np.pi) - np.pi def render(self, mode="human"): # if self.render_mode is not None: # return self.renderer.get_renders() # else: # return self._render(mode) return self._render(mode) def _blitRotate(self, surf, image, origin, pivot, angle): image_rect = image.get_rect(topleft = (origin[0] - pivot[0], origin[1]-pivot[1])) offset_center_to_pivot = pygame.math.Vector2(origin) - image_rect.center rotated_offset = offset_center_to_pivot.rotate(-angle) rotated_image_center = (origin[0] - rotated_offset.x, origin[1] - rotated_offset.y) rotated_image = pygame.transform.rotate(image, angle) rotated_image_rect = rotated_image.get_rect(center = rotated_image_center) surf.blit(rotated_image, rotated_image_rect) def get_pos_xy(self, manip_obj): return self.manip_objs[manip_obj]['pos_xy'] def get_pos_orientation(self, manip_obj): return self.manip_objs[manip_obj]['orientation'] def ik(self, xyo): def x_constraint(q, xyo): """Returns the corresponding hand xy coordinates for a given set of joint angle values [shoulder, elbow, wrist], and the above defined arm segment lengths, L q : np.array the list of current joint angles xy : np.array current xy position (not used) returns : np.array the difference between current and desired x position """ xy = xyo[:2] xy = [self.scale * x for x in xy] return self.lim_length * np.sin(q[0]) + self.lim_length * np.sin(q[0] + q[1]) + self.tool_center_point * np.sin(q[0] + q[1] + q[2]) - xy[0] def y_constraint(q, xyo): """Returns the corresponding hand xy coordinates for a given set of joint angle values [shoulder, elbow, wrist], and the above defined arm segment lengths, L q : np.array the list of current joint angles xy : np.array current xy position (not used) returns : np.array the difference between current and desired y position """ xy = xyo[:2] xy = [self.scale * x for x in xy] return self.lim_length * np.cos(q[0]) + self.lim_length * np.cos(q[0] + q[1]) + self.tool_center_point * np.cos(q[0] + q[1] + q[2]) - xy[1] def sin_o_constraint(q, xyo): """Returns the corresponding hand xy coordinates for a given set of joint angle values [shoulder, elbow, wrist], and the above defined arm segment lengths, L q : np.array the list of current joint angles xy : np.array current xy position (not used) returns : np.array the difference between current and desired y position """ o = xyo[-1] return (np.sin(q[0] + q[1] + q[2]) - np.sin(o)) ** 2 def cos_o_constraint(q, xyo): """Returns the corresponding hand xy coordinates for a given set of joint angle values [shoulder, elbow, wrist], and the above defined arm segment lengths, L q : np.array the list of current joint angles xy : np.array current xy position (not used) returns : np.array the difference between current and desired y position """ o = xyo[-1] return (np.cos(q[0] + q[1] + q[2]) - np.cos(o)) ** 2 + (np.sin(q[0] + q[1] + q[2]) - np.sin(o)) ** 2 def distance_to_default(q, *args): """Objective function to minimize Calculates the euclidean distance through joint space to the default arm configuration. The weight list allows the penalty of each joint being away from the resting position to be scaled differently, such that the arm tries to stay closer to resting state more for higher weighted joints than those with a lower weight. q : np.array the list of current joint angles returns : scalar euclidean distance to the default arm position """ # weights found with trial and error, # get some wrist bend, but not much weight = [1, 1, 0.5] try: action = np.sqrt(np.sum([(qi - q0i)**2 * wi for qi, q0i, wi in zip(q, self.robot_joints_init.tolist()[:-1], weight)])) except Exception: print(Exception) return action if len(xyo) == 3: ik_result = scipy.optimize.fmin_slsqp( func=distance_to_default, x0=self.robot_joints, eqcons=[x_constraint, y_constraint, # sin_o_constraint, cos_o_constraint], # uncomment to add in min / max angles for the joints # ieqcons=[joint_limits_upper_constraint, # joint_limits_lower_constraint], args=(xyo,), iprint=0) # iprint=0 suppresses output elif len(xyo) == 2: ik_result = scipy.optimize.fmin_slsqp( func=distance_to_default, x0=self.robot_joints, eqcons=[x_constraint, y_constraint, # sin_o_constraint, ], # uncomment to add in min / max angles for the joints # ieqcons=[joint_limits_upper_constraint, # joint_limits_lower_constraint], args=(xyo,), iprint=0) # iprint=0 suppresses output return ik_result def _calculate_img_starting_pos(self, img_pos, img_size): x = img_pos[0] - img_size[0] / 2 y = img_pos[1] - img_size[1] / 2 return [x, y] def _render(self, mode="human"): assert mode in self.metadata["render_modes"] try: import pygame from pygame import gfxdraw except ImportError: raise DependencyNotInstalled( "pygame is not installed, run `pip install gym[classic_control]`" ) if self.screen is None: pygame.init() if mode == "human": pygame.display.init() self.screen = pygame.display.set_mode( (self.screen_width, self.screen_height) ) else: # mode in {"rgb_array", "single_rgb_array"} self.screen = pygame.Surface((self.screen_width, self.screen_height)) if self.clock is None: self.clock = pygame.time.Clock() # world_width = 200 # scale = self.screen_width / world_width self.surf = pygame.Surface((self.screen_width, self.screen_height)) self.surf.fill((255, 255, 255)) self.surf.blit(self.env_objs['wood']['image'], (0, 0)) for obj in list(self.manip_objs.keys())[::-1]: if 'pos_xy' in self.manip_objs[obj]: try: image_transformed = pygame.transform.rotate( self.manip_objs[obj]['image'], np.rad2deg(self.manip_objs[obj]['orientation'])) except Exception as e: print(e) print(self.manip_objs[obj]['orientation'], np.rad2deg(self.manip_objs[obj]['orientation'])) # print(obj, self.manip_objs[obj]['orientation']) new_rect = image_transformed.get_rect() # self.surf.blit(image_lim_transformed, (mid_x - new_rect[2] / 2, mid_y - new_rect[3] / 2)) self.surf.blit(image_transformed, self._calculate_img_starting_pos( self.manip_objs[obj]['pos_xy'], [new_rect[2], new_rect[3]] )) # self.surf.blit(self.image_apple, self.positions[0] - self.size[0] / 2) # self.surf.blit(self.image_orange, self.positions[1] - self.size[1] / 2) # self.surf.blit(self.image_banana, self.positions[2] - self.size[2] / 2) start_x = 0 start_y = 0 end_x = self.robot_base_xy[0] end_y = self.robot_base_xy[1] angle = 0 for i in range(self.robot_joints.shape[0] - 1): if i < 2: start_x = end_x start_y = end_y angle = angle + self.robot_joints[i] end_x = start_x + np.sin(angle) * self.lim_length end_y = start_y + np.cos(angle) * self.lim_length mid_x = (start_x + end_x) / 2 mid_y = (start_y + end_y) / 2 image_lim_transformed = pygame.transform.rotate(self.env_objs['lim']['image'], np.rad2deg(angle)) new_rect = image_lim_transformed.get_rect() self.surf.blit(image_lim_transformed, (mid_x - new_rect[2] / 2, mid_y - new_rect[3] / 2)) elif i == 2: start_x = end_x start_y = end_y angle = angle + self.robot_joints[i] # print('gripper angle:', self.robot_joints[-1]) if self._gripper_closed_(): mid_x = start_x + np.sin(angle) * (self.env_objs['gripper_closed']['size_xy'][1] / 2 + 38 * self.scale) mid_y = start_y + np.cos(angle) * (self.env_objs['gripper_closed']['size_xy'][1] / 2 + 38 * self.scale) image_gripper_closed_transformed = pygame.transform.rotate(self.env_objs['gripper_closed']['image'], np.rad2deg(angle)) new_rect = image_gripper_closed_transformed.get_rect() self.surf.blit(image_gripper_closed_transformed, (mid_x - new_rect[2] / 2, mid_y - new_rect[3] / 2)) else: mid_x = start_x + np.sin(angle) * (self.env_objs['gripper_open']['size_xy'][1] / 2 + 38 * self.scale) mid_y = start_y + np.cos(angle) * (self.env_objs['gripper_open']['size_xy'][1] / 2 + 38 * self.scale) image_gripper_open_transformed = pygame.transform.rotate(self.env_objs['gripper_open']['image'], np.rad2deg(angle)) new_rect = image_gripper_open_transformed.get_rect() self.surf.blit(image_gripper_open_transformed, (mid_x - new_rect[2] / 2, mid_y - new_rect[3] / 2)) # self.surf = pygame.transform.flip(self.surf, False, True) self.screen.blit(self.surf, (0, 0)) if mode == "human": pygame.event.pump() self.clock.tick(self.metadata["render_fps"]) pygame.display.flip() elif mode in {"rgb_array", "single_rgb_array"}: return np.transpose( np.array(pygame.surfarray.pixels3d(self.screen)), axes=(1, 0, 2) ) def close(self): if self.screen is not None: import pygame pygame.display.quit() pygame.quit() self.isopen = False def get_base_xy(self): return self.robot_base_xy def get_joint_angles(self): return self.robot_joints if __name__ == '__main__': # env = TinyUR5Env(render_mode='rgb_array') env = TinyUR5Env(render_mode='human') observation, info = env.reset(seed=42, return_info=True) for i in range(1000): print(i) input() action = env.action_space.sample() observation, reward, done, info = env.step(action) img = env.render() # print(img[0].shape) if done: observation, info = env.reset(return_info=True) env.close()