from typing import Any, Dict, Union import numpy as np import sapien import torch from mani_skill.agents.robots import SO100, Fetch, Panda from mani_skill.envs.sapien_env import BaseEnv from mani_skill.envs.tasks.tabletop.pick_cube_cfgs import PICK_CUBE_CONFIGS from mani_skill.sensors.camera import CameraConfig from mani_skill.utils import sapien_utils from mani_skill.utils.building import actors from mani_skill.utils.registration import register_env from mani_skill.utils.scene_builder.table import TableSceneBuilder from mani_skill.utils.structs.pose import Pose from mani_skill.utils import common, sapien_utils from mani_skill.utils.structs import Actor #Robomme import matplotlib.pyplot as plt from mani_skill.utils.geometry.rotation_conversions import ( euler_angles_to_matrix, matrix_to_quaternion, ) from .utils import * from .utils.difficulty import normalize_robomme_difficulty from .utils.subgoal_evaluate_func import static_check from .utils import subgoal_language from .utils.object_generation import spawn_fixed_cube, build_board_with_hole from .utils import reset_panda from .utils import subgoal_evaluate_func from ..logging_utils import logger PICK_CUBE_DOC_STRING = """**Task Description:** A simple task where the objective is to grasp a red cube with the {robot_id} robot and move it to a target goal position. This is also the *baseline* task to test whether a robot with manipulation capabilities can be simulated and trained properly. Hence there is extra code for some robots to set them up properly in this environment as well as the table scene builder. **Randomizations:** - the cube's xy position is randomized on top of a table in the region [0.1, 0.1] x [-0.1, -0.1]. It is placed flat on the table - the cube's z-axis rotation is randomized to a random angle - the target goal position (marked by a green sphere) of the cube has its xy position randomized in the region [0.1, 0.1] x [-0.1, -0.1] and z randomized in [0, 0.3] **Success Conditions:** - the cube position is within `goal_thresh` (default 0.025m) euclidean distance of the goal position - the robot is static (q velocity < 0.2) """ @register_env("InsertPeg") class InsertPeg(BaseEnv): _sample_video_link = "https://github.com/haosulab/ManiSkill/raw/main/figures/environment_demos/PickCube-v1_rt.mp4" SUPPORTED_ROBOTS = [ "panda", "fetch", "xarm6_robotiq", "so100", "widowxai", ] agent: Union[Panda] goal_thresh = 0.025 cube_spawn_half_size = 0.05 cube_spawn_center = (0, 0) _clearance = 0.01 def __init__(self, *args, robot_uids="panda_wristcam", robot_init_qpos_noise=0,seed=0,Robomme_video_episode=None,Robomme_video_path=None, **kwargs): self.use_demonstrationwrapper=False self.demonstration_record_traj=False self.robot_init_qpos_noise = robot_init_qpos_noise if robot_uids in PICK_CUBE_CONFIGS: cfg = PICK_CUBE_CONFIGS[robot_uids] else: cfg = PICK_CUBE_CONFIGS["panda"] self.cube_half_size = cfg["cube_half_size"] self.goal_thresh = cfg["goal_thresh"] self.cube_spawn_half_size = cfg["cube_spawn_half_size"] self.cube_spawn_center = cfg["cube_spawn_center"] self.max_goal_height = cfg["max_goal_height"] self.sensor_cam_eye_pos = cfg["sensor_cam_eye_pos"] self.sensor_cam_target_pos = cfg["sensor_cam_target_pos"] self.human_cam_eye_pos = cfg["human_cam_eye_pos"] self.human_cam_target_pos = cfg["human_cam_target_pos"] self.seed = seed self._hb_generator = torch.Generator() self._hb_generator.manual_seed(int(self.seed)) self.robomme_failure_recovery = bool( kwargs.pop("robomme_failure_recovery", False) ) self.robomme_failure_recovery_mode = kwargs.pop( "robomme_failure_recovery_mode", None ) if isinstance(self.robomme_failure_recovery_mode, str): self.robomme_failure_recovery_mode = ( self.robomme_failure_recovery_mode.lower() ) normalized_robomme_difficulty = normalize_robomme_difficulty( kwargs.pop("difficulty", None) ) if normalized_robomme_difficulty is not None: self.difficulty = normalized_robomme_difficulty else: seed_mod = seed % 3 if seed_mod == 0: self.difficulty = "easy" elif seed_mod == 1: self.difficulty = "medium" else: self.difficulty = "hard" self.restore_flag=False super().__init__(*args, robot_uids=robot_uids, **kwargs) @property def _default_sensor_configs(self): pose = sapien_utils.look_at( eye=self.sensor_cam_eye_pos, target=self.sensor_cam_target_pos ) camera_eye=[0.3,0,0.4] camera_target =[0,0,-0.2] pose = sapien_utils.look_at( eye=camera_eye, target=camera_target ) return [CameraConfig("base_camera", pose, 256, 256, np.pi / 2, 0.01, 100)] @property def _default_human_render_camera_configs(self): pose = sapien_utils.look_at( eye=self.human_cam_eye_pos, target=self.human_cam_target_pos ) return CameraConfig("render_camera", pose, 512, 512, 1, 0.01, 100) def _load_agent(self, options: dict): super()._load_agent(options, sapien.Pose(p=[-0.615, 0, 0])) def _load_scene(self, options: dict): self.table_scene = TableSceneBuilder( self, robot_init_qpos_noise=0 ) self.table_scene.build() length_tensor = torch.rand(1, generator=self._hb_generator) radius_tensor = torch.rand(1, generator=self._hb_generator) # self.length = (0.05 + (0.0 - 0.0) * length_tensor).item() # self.radius = (0.01 + (0.0 - 0.0) * radius_tensor).item() self.length = (0.05 + (0.01 - 0.01) * length_tensor).item() self.radius = (0.01 + (0.005 - 0.005) * radius_tensor).item() # Create 3 identical pegs with different x-axis coordinates self.pegs = [] self.peg_heads = [] self.peg_tails = [] self._peg_initial_poses = [] offsets = [0.1,0,-0.1] # X-axis differences for the 3 pegs # Sample a single pair of colors so all pegs share the same appearance per seed. peg_head_color = torch.rand(3, generator=self._hb_generator).tolist() # Use complementary tail color so head/tail are contrasting. peg_tail_color = [1.0 - c for c in peg_head_color] for offset in offsets: peg_spawn_translation = np.array([self.length / 2 , -0.15-offset, self.radius], dtype=np.float32) #initial_yaw = (torch.rand(1, generator=self._hb_generator).item() * 2 * np.pi) - np.pi initial_yaw = 0 yaw_angles = torch.tensor([[0.0, 0.0, initial_yaw]], dtype=torch.float32) yaw_matrix = euler_angles_to_matrix(yaw_angles, convention="XYZ") yaw_quat = matrix_to_quaternion(yaw_matrix)[0].detach().cpu().numpy().tolist() peg_initial_pose = sapien.Pose( p=peg_spawn_translation.tolist(), q=yaw_quat, ) peg, peg_head, peg_tail = build_peg( self, length=self.length, radius=self.radius, initial_pose=peg_initial_pose, name=f"peg_{len(self.pegs)}", head_color=peg_head_color, tail_color=peg_tail_color, ) self.pegs.append(peg) self.peg_heads.append(peg_head) self.peg_tails.append(peg_tail) self._peg_initial_poses.append(peg_initial_pose) # Randomly select one peg from the 3 pegs random_peg_idx = int(torch.randint(0, 3, (1,), generator=self._hb_generator).item()) random_peg_idx=0 self.peg = self.pegs[random_peg_idx] self.peg_head = self.peg_heads[random_peg_idx] self.peg_tail = self.peg_tails[random_peg_idx] self._peg_initial_pose = self._peg_initial_poses[random_peg_idx] self.box=build_box_with_hole(self,inner_radius=self.radius*1.7,outer_radius=self.radius*4,depth=self.length,center=[0,0]) self.reset_in_proecess=False def _initialize_episode(self, env_idx: torch.Tensor, options: dict): with torch.device(self.device): self.end_steps=None self.table_scene.initialize(env_idx) # Reset highlight state at the start of each episode self._insert_highlight_start = None self._insert_highlight_active = False if not hasattr(self, "pegs"): return base_translation = (0, 0) x_jitter_2 = (torch.rand(1, generator=self._hb_generator).item() - 0.5) * 0.2 y_jitter_2 = (torch.rand(1, generator=self._hb_generator).item() - 0.5) * 0.2 # x_jitter_2=0 # y_jitter_2=0 box_translation = [base_translation[0] + x_jitter_2, base_translation[1] + y_jitter_2, self.radius * 4] box_yaw = np.pi / 2 + (torch.rand(1, generator=self._hb_generator).item() * 2 - 1) * np.radians(20) box_angles = torch.tensor([[0.0, 0.0, box_yaw]], dtype=torch.float32) box_matrix = euler_angles_to_matrix(box_angles, convention="XYZ") box_quat = matrix_to_quaternion(box_matrix)[0].detach().cpu().numpy().tolist() self.box.set_pose(sapien.Pose(p=box_translation, q=box_quat)) box_xy = np.array(box_translation[:2], dtype=np.float32) sampled_xy_positions = [] max_sampling_attempts = 512 #Initialize all 3 pegs with constrained random placements for i, peg in enumerate(self.pegs): candidate_xy = None for _ in range(max_sampling_attempts): x_sample = (torch.rand(1, generator=self._hb_generator).item() * 0.4) - 0.2 y_sample = (torch.rand(1, generator=self._hb_generator).item() * 0.6) - 0.3 sampled_xy = np.array([x_sample, y_sample], dtype=np.float32) if np.linalg.norm(sampled_xy - box_xy) <= self.radius * 6: continue if any(np.linalg.norm(sampled_xy - prev_xy) <= self.length * 1.5 for prev_xy in sampled_xy_positions): continue candidate_xy = sampled_xy break if candidate_xy is None: raise RuntimeError("Failed to sample peg positions satisfying placement constraints.") yaw_value = (torch.rand(1, generator=self._hb_generator).item() * 2 - 1) * np.radians(45) yaw_angles = torch.tensor([[0.0, 0.0, yaw_value]], dtype=torch.float32) yaw_matrix = euler_angles_to_matrix(yaw_angles, convention="XYZ") yaw_quat = matrix_to_quaternion(yaw_matrix)[0].detach().cpu().numpy().tolist() pose = sapien.Pose(p=[float(candidate_xy[0]), float(candidate_xy[1]), 0.0], q=yaw_quat) peg.set_pose(pose) sampled_xy_positions.append(candidate_xy) # Store initial poses for all pegs self.peg_init_poses = [] for peg in self.pegs: pose = peg.pose pose_p = np.asarray(pose.p, dtype=np.float32).reshape(-1).copy() pose_q = np.asarray(pose.q, dtype=np.float32).reshape(-1).copy() self.peg_init_poses.append(sapien.Pose(p=pose_p, q=pose_q)) # robomme-v2.7/robomme/robomme_env/PickPeg.py:243 pose = self.peg.pose pose_p = np.asarray(pose.p, dtype=np.float32).reshape(-1).copy() pose_q = np.asarray(pose.q, dtype=np.float32).reshape(-1).copy() self.peg_init_pose = sapien.Pose(p=pose_p, q=pose_q) self.peg_init_pose = sapien.Pose(p=pose_p, q=pose_q) # Define task list, each task contains a dictionary with function, name, demonstration flag, and optional failure_func obj_sample = torch.randint(0, 2, (1,), generator=self._hb_generator) dir_sample = torch.randint(0, 2, (1,), generator=self._hb_generator) self.obj_flag = -1 if obj_sample.item() == 0 else 1 self.direction = -1 if dir_sample.item() == 0 else 1 # if self.seed<30: # self.obj_flag=-1 # self.direction=1 # elif self.seed<60: # self.obj_flag=-1 # self.direction=-1 # elif self.seed<90: # self.obj_flag=1 # self.direction=1 # #elif self.seed<60: # else: # self.obj_flag=1 # self.direction=-1 # self.obj_flag=1 # self.direction=-1 qpos = np.array( [ 0.0, 0, 0, -np.pi * 4 / 8, 0, np.pi * 2 / 4, np.pi / 4, 0.04, 0.04, ], dtype=np.float32, ) self.agent.reset(qpos) if self.obj_flag==-1: self.grasp_target=self.peg_head self.insert_target=self.peg_tail else: self.grasp_target=self.peg_tail self.insert_target=self.peg_head agent_x = self.agent.robot.pose.p.tolist()[0][0] head_x = float(self.peg_head.pose.p.tolist()[0][0]) tail_x = float(self.peg_tail.pose.p.tolist()[0][0]) logger.debug(f"agent_x: {agent_x}, head_x: {head_x}, tail_x: {tail_x}") near_link = self.peg_head if abs(head_x - agent_x) <= abs(tail_x - agent_x) else self.peg_tail self.grasp_target_distance = "near" if self.grasp_target is near_link else "far" logger.debug(f"grasp_target_distance: {self.grasp_target_distance}") self.insert_way="left" if self.direction == -1 else "right" tasks = [ { "func": lambda: is_A_pickup_notB(self, self.grasp_target, self.insert_target), "name": f"Pick up the peg by grasping the {self.grasp_target_distance} end", "subgoal_segment":f"Pick up the peg by grasping the {self.grasp_target_distance} end at <>", "choice_label": "pick up the peg by grasping one end", "demonstration": True, "failure_func": lambda: is_A_pickup_notB(self, self.insert_target, self.grasp_target), "solve": lambda env, planner: grasp_and_lift_peg_side(env, planner, env.grasp_target), "segment":self.grasp_target }, { "func": lambda: is_A_insert_notB(self, self.insert_target, self.grasp_target, self.box,direction=self.direction), "name": f"Insert the peg from the {self.insert_way} side of the box", "subgoal_segment":f"Insert the peg from the {self.insert_way} side of the box at <>", "choice_label": f"insert the peg from the {self.insert_way} side", "demonstration": True, "failure_func": None, "solve": lambda env, planner: insert_peg(env, planner, direction=self.direction,obj=self.obj_flag,insert_obj=self.insert_target), "segment":self.box }, { "func": lambda:reset_check(self), "name": "NO RECORD", "subgoal_segment":f"NO RECORD", "demonstration": True, "failure_func": None, "specialflag":"reset pegs", "solve": lambda env, planner: [solve_strong_reset(env,planner)], }, { "func": lambda: static_check(self, timestep=int(self.elapsed_steps), static_steps=100), "name": "NO RECORD", "subgoal_segment":"NO RECORD", "demonstration": True, "failure_func": None, "solve": lambda env, planner: [solve_hold_obj(env, planner, static_steps=100,close=False)], }, { "func": lambda: is_A_pickup_notB(self, self.grasp_target, self.insert_target), "name": f"Pick up the peg by grasping the {self.grasp_target_distance} end", "subgoal_segment":f"Pick up the peg by grasping the {self.grasp_target_distance} end at <>", "choice_label": "pick up the peg by grasping one end", "demonstration": False, "failure_func": lambda: [ is_A_pickup_notB(self, self.insert_target, self.grasp_target), is_any_obj_pickup(self, [head for i, head in enumerate(self.peg_heads) if self.pegs[i] is not self.peg] + [tail for i, tail in enumerate(self.peg_tails) if self.pegs[i] is not self.peg]) ], "solve": lambda env, planner: grasp_and_lift_peg_side(env, planner, env.grasp_target), "segment":self.grasp_target }, { "func": lambda: is_A_insert_notB(self, self.insert_target, self.grasp_target,self.box,direction=self.direction,mark_end_flag=True), "name": f"Insert the peg from the {self.insert_way} side", "subgoal_segment":f"Insert the peg from the {self.insert_way} side at <>", "choice_label": f"insert the peg from the {self.insert_way} side", "demonstration": False, "failure_func": lambda: [ is_A_insert_notB(self, self.grasp_target, self.insert_target, self.box), is_A_insert_notB(self, self.insert_target, self.grasp_target, self.box, direction=-self.direction), is_any_obj_pickup(self, [head for i, head in enumerate(self.peg_heads) if self.pegs[i] is not self.peg] + [tail for i, tail in enumerate(self.peg_tails) if self.pegs[i] is not self.peg]) ], "solve": lambda env, planner: insert_peg(env, planner, direction=self.direction,obj=self.obj_flag,insert_obj=self.insert_target,cut_retreat=True), "segment":self.box }, ] # Store task list for RecordWrapper use self.task_list = tasks def evaluate(self,solve_complete_eval=False): timestep = self.elapsed_steps self.successflag=torch.tensor([False]) self.failureflag = torch.tensor([False]) # Use encapsulated sequence task check function if(self.use_demonstrationwrapper==False):# change subgoal after planner ends during recording if solve_complete_eval==True: allow_subgoal_change_this_timestep=True else: allow_subgoal_change_this_timestep=False else:# during demonstration, video needs to call evaluate(solve_complete_eval), video ends and flag changes in demonstrationwrapper if solve_complete_eval==True or self.demonstration_record_traj==False: allow_subgoal_change_this_timestep=True else: allow_subgoal_change_this_timestep=False all_tasks_completed, current_task_name, task_failed,_ = sequential_task_check(self, self.task_list,allow_subgoal_change_this_timestep=allow_subgoal_change_this_timestep) if self.end_steps!=None:# truncate tail, also truncate tail in planner logger.debug( "elapsed_steps=%s, end_steps=%s", self.elapsed_steps, self.end_steps, ) if int(getattr(self, "elapsed_steps", 0))>=self.end_steps+3: self.successflag = torch.tensor([True]) if task_failed: self.failureflag = torch.tensor([True]) logger.debug(f"Task failed: {current_task_name}") # If static_check succeeds or all tasks completed, set success flag if all_tasks_completed and not task_failed: self.successflag = torch.tensor([True]) return { "success": self.successflag, "fail": self.failureflag, } def compute_dense_reward(self, obs: Any, action: torch.Tensor, info: Dict): tcp_to_obj_dist = torch.linalg.norm( self.agent.tcp_pose.p - self.agent.tcp_pose.p, axis=1 ) reaching_reward = 1 - torch.tanh(5 * tcp_to_obj_dist) reward = reaching_reward*0 return reward def compute_normalized_dense_reward( self, obs: Any, action: torch.Tensor, info: Dict ): return self.compute_dense_reward(obs=obs, action=action, info=info) / 5 #Robomme def step(self, action: Union[None, np.ndarray, torch.Tensor, Dict]): obs, reward, terminated, truncated, info = super().step(action) cur_step = int(self.elapsed_steps[0].item()) if self.reset_in_proecess==True: for i, peg in enumerate(self.pegs): peg.set_pose(self.peg_init_poses[i]) if peg.dof > 0: zero = np.zeros(peg.dof) peg.set_qpos(zero) peg.set_qvel(zero) logger.debug("reset peg!") if is_A_insert_notB(self, self.insert_target, self.grasp_target, self.box,direction=self.direction): self.start_step=cur_step color=sapien.render.RenderMaterial( base_color=sapien_utils.hex2rgba("#FFD289"), roughness=0.5, specular=0.5) if getattr(self, "start_step", None) is not None : if cur_step <= self.start_step + 20 and cur_step>= self.start_step: color=[1.0, 0.0, 0.0, 1.0] highlight_obj( self, self.box, start_step= 0, end_step= 99999, cur_step=cur_step, disk_radius=0.015, disk_half_length=0.055, highlight_color=color,) return obs, reward, terminated, truncated, info