from typing import Any, Dict, Union import numpy as np import sapien import torch import mani_skill.envs.utils.randomization as randomization 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 #Robomme import matplotlib.pyplot as plt import random from mani_skill.utils.geometry.rotation_conversions import ( euler_angles_to_matrix, matrix_to_quaternion, ) from .utils import * 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.difficulty import normalize_robomme_difficulty 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("PickHighlight") class PickHighlight(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) config_hard = { 'spawn': 6, "pickup": 3 } config_easy = { 'spawn': 3, "pickup": 1 } config_medium = { 'spawn': 4, "pickup": 2 } # Combine into a dictionary configs = { 'hard': config_hard, 'easy': config_easy, 'medium': config_medium } 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.robot_init_qpos_noise = robot_init_qpos_noise self.use_demonstrationwrapper=False self.demonstration_record_traj=False 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.generator = torch.Generator() self.generator.manual_seed(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: # Determine difficulty based on seed % 3 seed_mod = seed % 3 if seed_mod == 0: self.difficulty = "easy" elif seed_mod == 1: self.difficulty = "medium" else: # seed_mod == 2 self.difficulty = "hard" 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.generator.manual_seed(self.seed) self.table_scene = TableSceneBuilder( self, robot_init_qpos_noise=self.robot_init_qpos_noise ) self.table_scene.build() button_obb = build_button( self, center_xy=(-0.2, 0), scale=1.5, generator=self.generator, ) avoid = [button_obb] self.all_cubes = [] # Save all cube objects self.all_cube_names = [] self.all_cube_colors = [] # List of available colors available_colors = [ {"color": (1, 0, 0, 1), "name": "red"}, {"color": (0, 0, 1, 1), "name": "blue"}, {"color": (0, 1, 0, 1), "name": "green"} ] # Get number of cubes to spawn based on difficulty num_cubes_to_spawn = self.configs[self.difficulty]['spawn'] # Spawn specified number of cubes, each with random color for cube_idx in range(num_cubes_to_spawn): # Randomly select a color color_choice_idx = torch.randint(0, len(available_colors), (1,), generator=self.generator).item() chosen_color = available_colors[color_choice_idx] try: cube = spawn_random_cube( self, color=chosen_color["color"], avoid=avoid, include_existing=False, include_goal=False, region_center=[-0.1, 0], region_half_size=0.2, half_size=self.cube_half_size, min_gap=self.cube_half_size*2, random_yaw=True, name_prefix=f"cube_{chosen_color['name']}_{cube_idx}", generator=self.generator, ) cube_name = f"cube_{chosen_color['name']}_{cube_idx}" # Add cube immediately after successful creation self.all_cubes.append(cube) self.all_cube_names.append(cube_name) self.all_cube_colors.append(chosen_color["name"]) setattr(self, cube_name, cube) avoid.append(cube) except RuntimeError as e: logger.debug(f"Failed to spawn cube {cube_idx} ({chosen_color['name']}): {e}") break logger.debug(f"Generated {len(self.all_cubes)} cubes total") # Randomly select one cube from all available cubes as the target target_cube_indices = torch.randperm(len(self.all_cubes), generator=self.generator)[:self.configs[self.difficulty]['pickup']] self.target_cubes = [self.all_cubes[idx] for idx in target_cube_indices] self.target_cube_names = [self.all_cube_names[idx] for idx in target_cube_indices] self.target_cube_colors = [self.all_cube_colors[idx] for idx in target_cube_indices] self.target_labels = [ (color or name or "target") for color, name in zip(self.target_cube_colors, self.target_cube_names) ] # Record pick count for each target cube; all must be > 1 for success self.target_cube_pickup_counts = {name: 0 for name in self.target_cube_names} # Define task list, each task contains a dictionary with function, name, demonstration flag, and optional failure_func tasks = [] target_label = getattr(self, "target_cube_color", None) or getattr( self, "target_cube_name", None ) or getattr(self, "target_label", None) or "target" self.target_label = target_label tasks.append({ "func": lambda: is_button_pressed(self, obj=self.button), "name": "press the button", "subgoal_segment":"press the button at <>", "choice_label": "press button", "demonstration": False, "failure_func":is_any_obj_pickup(self,[cube for cube in self.all_cubes]), "solve": lambda env, planner:solve_button(env, planner, obj=self.button), "segment":self.cap_link, }) # Pick each target cube once, lambda captures current cube explicitly to avoid closure issue num_targets = len(self.target_cubes) for cube_idx, cube in enumerate(self.target_cubes): # If only one target cube, do not show index if num_targets == 1: task_name = f"pick up the highlighted cube, which is {self.target_labels[cube_idx]}" task_subgoal = f"pick up the highlighted cube at <>, which is {self.target_labels[cube_idx]}" else: task_name = subgoal_language.get_subgoal_with_index(cube_idx, "pick up the {idx} highlighted cube, which is {color}", color=self.target_labels[cube_idx]) task_subgoal = subgoal_language.get_subgoal_with_index(cube_idx, "pick up the {idx} highlighted cube at <>, which is {color}", color=self.target_labels[cube_idx]) tasks.append({ "func": (lambda c=cube: is_any_obj_pickup_flag_currentpickup(self, objects=[c])), "name": task_name, "subgoal_segment": task_subgoal, "choice_label": "pick up the highlighted cube", "demonstration": False, "failure_func": lambda idx=cube_idx: [is_any_obj_pickup(self,[cube for cube in self.all_cubes if cube not in self.target_cubes] ), ], "solve": lambda env, planner, c=cube: solve_pickup(env, planner, obj=c), "segment":cube, }) if cube_idx!=num_targets-1: tasks.append({ "func": (lambda :is_obj_dropped_currentpickup(self,self.target_cubes)), "name": f"place the cube onto the table", "subgoal_segment":"place the cube onto the table", "choice_label": "place the cube onto the table", "demonstration": False, "failure_func": lambda idx=cube_idx: [ is_any_obj_pickup(self,[cube for cube in self.all_cubes if cube not in self.target_cubes] ), ], "solve": lambda env, planner, c=cube: [solve_putdown_whenhold(env, planner, release_z=0.01), # solve_pickup(env, planner, obj=c), # solve_putdown_whenhold(env, planner, obj=c,release_z=0.01)# For testing ], "segment":None, }) # Store task list for RecordWrapper use self.task_list = tasks # Record pickup related task indices and items for recovery self.recovery_pickup_indices, self.recovery_pickup_tasks = task4recovery(self.task_list) if self.robomme_failure_recovery: # Only inject an intentional failed grasp when recovery mode is enabled self.fail_grasp_task_index = inject_fail_grasp( self.task_list, generator=self.generator, mode=self.robomme_failure_recovery_mode, ) else: self.fail_grasp_task_index = None def _initialize_episode(self, env_idx: torch.Tensor, options: dict): with torch.device(self.device): b = len(env_idx) self.table_scene.initialize(env_idx) qpos=reset_panda.get_reset_panda_param("qpos") self.agent.reset(qpos) def _get_obs_extra(self, info: Dict): return dict() def evaluate(self,solve_complete_eval=False): self.successflag=torch.tensor([False]) # Keep previous failure state (once failed, always failed) if not hasattr(self, 'failureflag') or self.failureflag is None: self.failureflag = torch.tensor([False]) previous_failure = bool(self.failureflag.detach().cpu().item()) if isinstance(self.failureflag, torch.Tensor) else False # If previously failed, do not reset, keep failed state; otherwise reset if previous_failure: # Keep failed state, do not reset pass else: self.failureflag = torch.tensor([False]) 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 # Use encapsulated sequence task check function all_tasks_completed, current_task_name, task_failed ,self.current_task_specialflag= sequential_task_check(self, self.task_list,allow_subgoal_change_this_timestep=allow_subgoal_change_this_timestep) if task_failed: self.failureflag = torch.tensor([True]) logger.debug(f"Task failed: {current_task_name}") # If previously failed, keep failed state if previous_failure: self.failureflag = torch.tensor([True]) ############# Rising edge detection must be placed before fail detection target_cubes = getattr(self, "target_cubes", []) target_cube_names = getattr(self, "target_cube_names", []) if target_cubes and not hasattr(self, "target_cube_pickup_counts"): self.target_cube_pickup_counts = {name: 0 for name in target_cube_names} self.target_cube_pickup_active = {name: False for name in target_cube_names} if target_cubes and not hasattr(self, "target_cube_pickup_active"): self.target_cube_pickup_active = {name: False for name in target_cube_names} # Only count when cube changes from "not picked" to "picked", avoid duplicate counting in multiple frames for same pick for cube, name in zip(target_cubes, target_cube_names): pickup_tensor = is_obj_pickup(self, cube) if isinstance(pickup_tensor, torch.Tensor): picked_now = bool(pickup_tensor.detach().cpu().any()) else: picked_now = bool(pickup_tensor) was_picked = self.target_cube_pickup_active.get(name, False) if picked_now and not was_picked: self.target_cube_pickup_counts[name] = ( self.target_cube_pickup_counts.get(name, 0) + 1 ) self.target_cube_pickup_active[name] = picked_now pickup_counts = getattr(self, "target_cube_pickup_counts", {}) counts_satisfied = ( len(pickup_counts) > 0 and all(count >= 1 for count in pickup_counts.values()) ) ############# Rising edge detection must be placed before fail detection # Success if all picked at least once (counting discrete pick events) if counts_satisfied: self.successflag = torch.tensor([True]) # Fail if planner finished but not successful if all_tasks_completed and not counts_satisfied: self.failureflag = torch.tensor([True]) logger.debug(f"Pickup counts not satisfied: {pickup_counts}") if self.failureflag == torch.tensor([True]): pass 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]): timestep = self.elapsed_steps target_cubes = getattr(self, "target_cubes", []) highlight_count = min(self.configs[self.difficulty]["pickup"], len(target_cubes)) for i in range(highlight_count): highlight_obj( self, target_cubes[i], start_step=10, end_step=100, cur_step=timestep, ) obs, reward, terminated, truncated, info = super().step(action) return obs, reward, terminated, truncated, info