"""Rollout config for converted scenario output. Each object is spawned individually from a shared USD model library via ``UsdFileCfg``, giving Isaac Lab full control over spawning and physics. The station is spawned as a static ``AssetBaseCfg``:: cfg = ScenarioIKRolloutCfg() cfg.dynamic_setup( scene_path="./configs/3_cabot_breakfast/spawn_config.json", library_dir="/path/to/usd_library", ) """ from __future__ import annotations import copy import json from pathlib import Path import isaaclab.sim as sim_utils from isaaclab.assets import ArticulationCfg, AssetBaseCfg, RigidObjectCfg from isaaclab.envs import ManagerBasedRLEnvCfg, ViewerCfg from isaaclab.managers import EventTermCfg as EventTerm from isaaclab.managers import ObservationGroupCfg as ObsGroup from isaaclab.managers import ObservationTermCfg as ObsTerm from isaaclab.managers import RewardTermCfg as RewTerm from isaaclab.managers import SceneEntityCfg from isaaclab.managers import TerminationTermCfg as DoneTerm from isaaclab.scene import InteractiveSceneCfg from isaaclab.sensors import ContactSensorCfg, TiledCameraCfg from isaaclab.utils import configclass import sim_improvement.environments.mdp as mdp from sim_improvement.environments.lbm.robot import ( IMPLICIT_PANDA, GRIPPER_ACTION, BINARY_GRIPPER_ACTION, LEFT_PANDA_DEFAULT_JOINT_POS, # LBM_GRAVITY_COMP_DIFF_IK, ARM_IK_ACTION, DUAL_ARM_IK_ACTION, RIGHT_PANDA_DEFAULT_JOINT_POS, ) from sim_improvement.environments.lbm.scenario_helper import ( ResetScenarioStochastic, YUP_OBJECTS_REGISTRY, _compute_child_frame_z_offset, _quat_mul_single, _midpoint_translation, _midpoint_rotation, _RX90_W, _RX90_X, ) # --------------------------------------------------------------------------- # Scene # --------------------------------------------------------------------------- @configclass class ScenarioSceneCfg(InteractiveSceneCfg): """Scene config that dynamically adds objects from converter output.""" left_panda = IMPLICIT_PANDA.replace( prim_path="{ENV_REGEX_NS}/left_panda", init_state=ArticulationCfg.InitialStateCfg( pos=(-0.5937, -0.34362, -0.08484), rot=(0.36811, 0.01027, 0.00078, 0.92973), joint_pos=LEFT_PANDA_DEFAULT_JOINT_POS, ), ) right_panda = IMPLICIT_PANDA.replace( prim_path="{ENV_REGEX_NS}/right_panda", init_state=ArticulationCfg.InitialStateCfg( pos=(-0.5937, 0.32962, -0.08062), rot=(0.91675, 0.01312, 0.01089, 0.39909), joint_pos=RIGHT_PANDA_DEFAULT_JOINT_POS, ), ) background = AssetBaseCfg( prim_path="{ENV_REGEX_NS}/background", spawn=sim_utils.DomeLightCfg( intensity=500.0, texture_file = "envs/2_riverway_shelf/poly_haven_studio_4k.hdr" ), init_state=AssetBaseCfg.InitialStateCfg( pos=(0.0, 0.0, 0.0), rot=(0.0, 0.0, 0.0, 1.0), ) ) # Placeholder — filled by dynamic_setup scenario: AssetBaseCfg | None = None def dynamic_setup(self, spawn_config: dict, library_dir: str, scene_path_key: str) -> None: """Parse converter output and populate scene entities. Args: spawn_config: Parsed spawn_config.json dict. library_dir: Path to shared USD model library. scene_path_key: Resolved path string used as key for YUP_OBJECTS_REGISTRY. """ library_path = Path(library_dir) yup_objects: set[str] = set() YUP_OBJECTS_REGISTRY[scene_path_key] = yup_objects # Collect bodies that need contact reporting from predicates contact_bodies: set[str] = set() for pred in spawn_config.get("task_predicate_configs", []): if pred.get("_type") == "contactflagpredicateconfig": contact_bodies.add(pred["body_1_name"]) contact_bodies.add(pred["body_2_name"]) # 1. Spawn station as static asset station = spawn_config.get("station") if station and station.get("model"): station_usd = str(library_path / station["model"]) self.scenario = AssetBaseCfg( prim_path="{ENV_REGEX_NS}/scenario", spawn=sim_utils.UsdFileCfg(usd_path=station_usd), ) # 2. Spawn each manipulable object with UsdFileCfg spawned_objects: set[str] = set() for obj in spawn_config.get("objects", []): name = obj["name"] model_path = obj.get("model") if not model_path: continue usd_path = str(library_path / model_path) # Compute midpoint of stochastic translation for init_state pos = _midpoint_translation(obj.get("translation")) orient = _midpoint_rotation(obj.get("rotation")) # Track GLTF-sourced models that need Y-up correction at reset needs_yup = bool(obj.get("needs_yup")) if needs_yup: yup_objects.add(name) rx90 = (_RX90_W, _RX90_X, 0.0, 0.0) orient = _quat_mul_single(orient, rx90) # Lift objects whose child_frame is "bottom" so they sit on the surface z_off = _compute_child_frame_z_offset(usd_path, obj.get("child_frame"), needs_yup) pos = (pos[0], pos[1], pos[2] + z_off) # Welded objects are fixed in place → spawn as kinematic. is_kinematic = bool(obj.get("welded", False)) or bool(obj.get("kinematic_enabled", False)) no_gravity = is_kinematic or bool(obj.get("disable_gravity", False)) rigid_props = None if is_kinematic or no_gravity: rigid_props = sim_utils.RigidBodyPropertiesCfg( kinematic_enabled=is_kinematic, disable_gravity=no_gravity, ) asset = RigidObjectCfg( prim_path=f"{{ENV_REGEX_NS}}/{name}", spawn=sim_utils.UsdFileCfg( usd_path=usd_path, activate_contact_sensors=True, rigid_props=rigid_props, ), init_state=RigidObjectCfg.InitialStateCfg( pos=pos, rot=orient, ), ) setattr(self, name, asset) spawned_objects.add(name) # 3. Spawn uniform_manipuland groups (first variant) for group in spawn_config.get("randomized_groups", []): if group.get("group_type") != "uniform_manipuland": continue name = group["name"] variants = group.get("variants", []) if not variants: continue first = variants[0] model_path = first.get("model") if not model_path: continue usd_path = str(library_path / model_path) pos = _midpoint_translation(first.get("translation")) orient = _midpoint_rotation(first.get("rotation")) # Track GLTF-sourced models that need Y-up correction at reset needs_yup = bool(first.get("needs_yup")) if needs_yup: yup_objects.add(name) rx90 = (_RX90_W, _RX90_X, 0.0, 0.0) orient = _quat_mul_single(orient, rx90) z_off = _compute_child_frame_z_offset(usd_path, first.get("child_frame"), needs_yup) pos = (pos[0], pos[1], pos[2] + z_off) is_kinematic = bool(first.get("welded", False)) or bool(first.get("kinematic_enabled", False)) no_gravity = is_kinematic or bool(first.get("disable_gravity", False)) rigid_props = None if is_kinematic or no_gravity: rigid_props = sim_utils.RigidBodyPropertiesCfg( kinematic_enabled=is_kinematic, disable_gravity=no_gravity, ) asset = RigidObjectCfg( prim_path=f"{{ENV_REGEX_NS}}/{name}", spawn=sim_utils.UsdFileCfg( usd_path=usd_path, activate_contact_sensors=True, rigid_props=rigid_props, ), init_state=RigidObjectCfg.InitialStateCfg( pos=pos, rot=orient, ), ) setattr(self, name, asset) spawned_objects.add(name) # 4. Add ContactSensors for contact-flag predicates for pred in spawn_config.get("task_predicate_configs", []): if pred.get("_type") != "contactflagpredicateconfig": continue b1 = pred["body_1_name"] b2 = pred["body_2_name"] # body1 must be a spawned object (sensor prim_path). # body2 can be a spawned object OR a station-embedded prim. if b1 not in spawned_objects: continue if b2 in spawned_objects: filter_path = f"{{ENV_REGEX_NS}}/{b2}" else: # Assume station-embedded: {ENV_REGEX_NS}/scenario/{body_name} filter_path = f"{{ENV_REGEX_NS}}/scenario/{b2}" sensor_name = f"contact__{b1}__{b2}" sensor = ContactSensorCfg( prim_path=f"{{ENV_REGEX_NS}}/{b1}", filter_prim_paths_expr=[filter_path], update_period=0.0, ) setattr(self, sensor_name, sensor) # --------------------------------------------------------------------------- # Actions # --------------------------------------------------------------------------- @configclass class ActionsCfg: # left_panda_arm = ARM_IK_ACTION.replace(asset_name="left_panda") # right_panda_arm = ARM_IK_ACTION.replace(asset_name="right_panda") dual_arm_ik_action = DUAL_ARM_IK_ACTION # left_panda_gripper = GRIPPER_ACTION.replace(asset_name="left_panda") # right_panda_gripper = GRIPPER_ACTION.replace(asset_name="right_panda") left_panda_gripper = BINARY_GRIPPER_ACTION.replace(asset_name="left_panda") right_panda_gripper = BINARY_GRIPPER_ACTION.replace(asset_name="right_panda") # --------------------------------------------------------------------------- # Observations # --------------------------------------------------------------------------- @configclass class VisionCfg(ObsGroup): wrist_camera_right = ObsTerm( func=mdp.image, params={"sensor_cfg": SceneEntityCfg("flir_wrist_right_minus"), "data_type": "rgb", "normalize": False}, ) wrist_camera_left = ObsTerm( func=mdp.image, # params={"sensor_cfg": SceneEntityCfg("wrist_camera_left"), "data_type": "rgb", "normalize": False}, params={"sensor_cfg": SceneEntityCfg("flir_wrist_left_plus"), "data_type": "rgb", "normalize": False}, ) external_camera_right = ObsTerm( func=mdp.image, params={"sensor_cfg": SceneEntityCfg("scene_camera_right"), "data_type": "rgb", "normalize": False}, ) external_camera_left = ObsTerm( func=mdp.image, params={"sensor_cfg": SceneEntityCfg("scene_camera_left"), "data_type": "rgb", "normalize": False}, ) left_ee_pos = ObsTerm( func=mdp.link_pos, params={"root_asset_cfg": SceneEntityCfg("left_panda"), "link_name": "panda_link8"}, ) left_ee_quat = ObsTerm( func=mdp.link_quat, params={"root_asset_cfg": SceneEntityCfg("left_panda"), "link_name": "panda_link8"}, ) right_ee_pos = ObsTerm( func=mdp.link_pos, params={"root_asset_cfg": SceneEntityCfg("right_panda"), "link_name": "panda_link8"}, ) right_ee_quat = ObsTerm( func=mdp.link_quat, params={"root_asset_cfg": SceneEntityCfg("right_panda"), "link_name": "panda_link8"}, ) left_gripper_pos = ObsTerm( func=lambda env, asset_cfg: mdp.joint_pos(env, asset_cfg) * -2, params={"asset_cfg": SceneEntityCfg(name="left_panda", joint_names=["panda_finger_joint1"])}, ) right_gripper_pos = ObsTerm( func=lambda env, asset_cfg: mdp.joint_pos(env, asset_cfg) * -2, params={"asset_cfg": SceneEntityCfg(name="right_panda", joint_names=["panda_finger_joint1"])}, ) def __post_init__(self): self.concatenate_terms = False self.enable_corruption = False @configclass class PolicyCfg(ObsGroup): """Observations for policy group.""" # prev_actions = ObsTerm(func=mdp.last_action) left_panda_joint_pos = ObsTerm(func=mdp.joint_pos, params={"asset_cfg": SceneEntityCfg("left_panda")}) right_panda_joint_pos = ObsTerm(func=mdp.joint_pos, params={"asset_cfg": SceneEntityCfg("right_panda")}) left_ee_pos = ObsTerm( func=mdp.link_pos, params={"root_asset_cfg": SceneEntityCfg("left_panda"), "link_name": "panda_link8"}, ) left_ee_quat = ObsTerm( func=mdp.link_quat, params={"root_asset_cfg": SceneEntityCfg("left_panda"), "link_name": "panda_link8"}, ) right_ee_pos = ObsTerm( func=mdp.link_pos, params={"root_asset_cfg": SceneEntityCfg("right_panda"), "link_name": "panda_link8"}, ) right_ee_quat = ObsTerm( func=mdp.link_quat, params={"root_asset_cfg": SceneEntityCfg("right_panda"), "link_name": "panda_link8"}, ) left_gripper_pos = ObsTerm( func=lambda env, asset_cfg: mdp.joint_pos(env, asset_cfg) * -2, params={"asset_cfg": SceneEntityCfg(name="left_panda", joint_names=["panda_finger_joint1"])}, ) right_gripper_pos = ObsTerm( func=lambda env, asset_cfg: mdp.joint_pos(env, asset_cfg) * -2, params={"asset_cfg": SceneEntityCfg(name="right_panda", joint_names=["panda_finger_joint1"])}, ) @configclass class ObservationsCfg: policy: PolicyCfg = PolicyCfg(concatenate_terms=True, enable_corruption=False) def dynamic_setup(self, spawn_config: dict) -> None: """Add task-relevant object pose obs to PolicyCfg from spawn config.""" # Build set of actual spawned scene entities to filter against scene_entities: set[str] = set() for obj in spawn_config.get("objects", []): if obj.get("model"): scene_entities.add(obj["name"]) for group in spawn_config.get("randomized_groups", []): if group.get("group_type") == "uniform_manipuland" and group.get("variants"): scene_entities.add(group["name"]) obj_names: set[str] = set() for pred in spawn_config.get("task_predicate_configs", []): ptype = pred.get("_type", "") if ptype == "ontoppredicateconfig": obj_names.add(pred["top_object_name"]) obj_names.add(pred["bottom_object_name"]) elif ptype == "contactflagpredicateconfig": obj_names.add(pred["body_1_name"]) for rc in spawn_config.get("reward_configs", []): for key in ("frame_a_name", "frame_b_name"): name = rc.get(key, "").split("::")[0] if name: obj_names.add(name) for name in spawn_config.get("observation_objects", []): obj_names.add(name) # Only add observations for names that are actual scene entities # (skip reference_frames and other non-spawned names) obj_names &= scene_entities for obj_name in sorted(obj_names): setattr(self.policy, f"{obj_name}_pos", ObsTerm( func=mdp.object_pos, params={"asset_cfg": SceneEntityCfg(obj_name)}, )) setattr(self.policy, f"{obj_name}_quat", ObsTerm( func=mdp.object_quat, params={"asset_cfg": SceneEntityCfg(obj_name)}, )) self.policy_dict = copy.deepcopy(self.policy) self.policy_dict.concatenate_terms = False self.policy_dict.enable_corruption = False # --------------------------------------------------------------------------- # Rewards / Terminations # --------------------------------------------------------------------------- @configclass class RewardsCfg: # translation_distance is added dynamically when reward_configs are present. success = RewTerm(func=mdp.success_reward, weight=10.0) # left_singularity = RewTerm( # func=mdp.singularity_penalty, weight=-1.0, # params={"asset_cfg": SceneEntityCfg("left_panda"), "threshold": 0.05}, # ) # right_singularity = RewTerm( # func=mdp.singularity_penalty, weight=-1.0, # params={"asset_cfg": SceneEntityCfg("right_panda"), "threshold": 0.05}, # ) def dynamic_setup(self, spawn_config: dict) -> None: """Add reward terms from spawn_config reward_configs.""" # Build set of actual spawned scene entities scene_entities: set[str] = set() for obj in spawn_config.get("objects", []): if obj.get("model"): scene_entities.add(obj["name"]) for group in spawn_config.get("randomized_groups", []): if group.get("group_type") == "uniform_manipuland" and group.get("variants"): scene_entities.add(group["name"]) for rc in spawn_config.get("reward_configs", []): if rc.get("_type") == "relativetranslationdistancecostconfig": obj_a = rc.get("frame_a_name", "").split("::")[0] obj_b = rc.get("frame_b_name", "").split("::")[0] if obj_a and obj_b and obj_a in scene_entities and obj_b in scene_entities: setattr(self, f"obj_dist__{obj_a}__{obj_b}", RewTerm( func=mdp.object_to_object_distance, weight=1 / 450.0, params={ "std": 0.3, "object1_cfg": SceneEntityCfg(obj_a), "object2_cfg": SceneEntityCfg(obj_b), }, )) setattr(self, f"ee_reach__{obj_b}", RewTerm( func=mdp.ee_to_object_distance, weight=1 / 450.0, params={ "std": 0.3, "object_cfg": SceneEntityCfg(obj_b), }, )) @configclass class TerminationsCfg: timeout = DoneTerm(func=mdp.time_out, time_out=True) left_singularity = DoneTerm( func=mdp.singularity_termination, params={"asset_cfg": SceneEntityCfg("left_panda"), "threshold": 0.02}, ) right_singularity = DoneTerm( func=mdp.singularity_termination, params={"asset_cfg": SceneEntityCfg("right_panda"), "threshold": 0.02}, ) # task_success is added dynamically when task_predicate_configs are present. def dynamic_setup(self, spawn_config: dict) -> None: """Add termination terms from spawn_config task_predicate_configs.""" predicates = spawn_config.get("task_predicate_configs", []) if predicates: # Build frame lookup from objects' "frames" dicts object_frames: dict[str, dict] = {} yup_objects: set[str] = set() for obj in spawn_config.get("objects", []): if "frames" in obj: object_frames[obj["name"]] = obj["frames"] if obj.get("needs_yup"): yup_objects.add(obj["name"]) self.success = DoneTerm( func=mdp.scenario_success, params={"predicates": predicates, "object_frames": object_frames, "yup_objects": yup_objects}, ) # --------------------------------------------------------------------------- # Events # --------------------------------------------------------------------------- @configclass class ScenarioEventCfg: reset_default = EventTerm( func=mdp.reset_scene_to_default, mode="reset", params={"reset_joint_targets": True}, ) reset_objects = EventTerm( func=ResetScenarioStochastic, mode="reset", params={"scene_path": ""}, # filled by dynamic_setup ) def dynamic_setup(self, scene_path: str, library_dir: str) -> None: """Set resolved paths for the stochastic reset event.""" if hasattr(self, "reset_objects"): self.reset_objects.params["scene_path"] = scene_path self.reset_objects.params["library_dir"] = library_dir # --------------------------------------------------------------------------- # Env # --------------------------------------------------------------------------- @configclass class ScenarioEnvCfg(ManagerBasedRLEnvCfg): scene = ScenarioSceneCfg(num_envs=32, env_spacing=5.0) observations = ObservationsCfg() rewards = RewardsCfg() terminations = TerminationsCfg() actions = ActionsCfg() events = ScenarioEventCfg() viewer: ViewerCfg = ViewerCfg(eye=(5.0, 0, 2.0), origin_type="world", env_index=0, asset_name="scenario") def __post_init__(self): self.seed = 42 self.decimation = 8 # self.episode_length_s = 45.0 self.episode_length_s = 30.0 self.sim.dt = 1 / (10.0 * self.decimation) self.sim.physx.solver_type = 1 self.sim.physx.max_position_iteration_count = 192 self.sim.physx.max_velocity_iteration_count = 1 self.sim.physx.bounce_threshold_velocity = 0.02 self.sim.physx.friction_offset_threshold = 0.01 self.sim.physx.friction_correlation_distance = 0.0005 self.sim.physx.gpu_found_lost_aggregate_pairs_capacity = 1024 * 1024 * 4 self.sim.physx.gpu_total_aggregate_pairs_capacity = 2**23 self.sim.physx.gpu_max_rigid_contact_count = 2**23 self.sim.physx.gpu_max_rigid_patch_count = 2**23 self.sim.physx.gpu_collision_stack_size = 2**31 self.sim.render.enable_dlssg = True self.sim.render.enable_ambient_occlusion = True self.sim.render.enable_reflections = True self.sim.render.enable_dl_denoiser = True self.wait_for_textures = True self.rerender_on_reset = True def dynamic_setup(self, scene_path: str, library_dir: str) -> None: """Configure the env from converter output directory. Loads spawn_config.json once and delegates to each sub-config's ``dynamic_setup`` so every manager owns its own wiring logic. Args: scene_path: Path to the config JSON file (spawn_config.json). library_dir: Path to shared USD model library. """ resolved_path = str(Path(scene_path).resolve()) resolved_lib = str(Path(library_dir).resolve()) with open(resolved_path) as f: spawn_config = json.load(f) self.scene.dynamic_setup(spawn_config, resolved_lib, scene_path_key=resolved_path) self.events.dynamic_setup(resolved_path, resolved_lib) self.terminations.dynamic_setup(spawn_config) self.rewards.dynamic_setup(spawn_config) self.observations.dynamic_setup(spawn_config) @configclass class ScenarioIKStateCfg(ScenarioEnvCfg): def __post_init__(self): super().__post_init__() self.actions.dual_arm_ik_action.controller.use_relative_mode = False self.scene.left_panda.spawn.rigid_props.disable_gravity = True self.scene.right_panda.spawn.rigid_props.disable_gravity = True @configclass class ScenarioIkVisionCfg(ScenarioIKStateCfg): def __post_init__(self): super().__post_init__() # Cameras — same as rollout_cfg.py self.scene.scene_camera_right = TiledCameraCfg( prim_path="{ENV_REGEX_NS}/scene_camera_right", offset=TiledCameraCfg.OffsetCfg( pos=(0.54755, 0.18739, 0.83597), rot=(0.63083, 0.15475, 0.27309, 0.70960), convention="opengl", ), data_types=["rgb"], spawn=sim_utils.PinholeCameraCfg( clipping_range=(0.05, 3.0), focal_length=0.5953, horizontal_aperture=1.0, vertical_aperture=0.75, ), height=480, width=640, ) self.scene.scene_camera_left = TiledCameraCfg( prim_path="{ENV_REGEX_NS}/scene_camera_left", offset=TiledCameraCfg.OffsetCfg( pos=(0.49714, -0.36177, 0.77877), rot=(0.72133, 0.33411, 0.07049, 0.60257), convention="opengl", ), data_types=["rgb"], spawn=sim_utils.PinholeCameraCfg( clipping_range=(0.05, 3.0), focal_length=0.5953, horizontal_aperture=1.0, vertical_aperture=0.75, ), height=480, width=640, ) self.scene.flir_wrist_left_plus = TiledCameraCfg( prim_path="{ENV_REGEX_NS}/left_panda/panda_link8/flir_wrist_left_plus", offset=TiledCameraCfg.OffsetCfg( pos=(0.04651, -0.04651, 0.0048), rot=(0.92388, 0.0, 0.0, 0.38268), convention="ros", ), data_types=["rgb"], spawn=sim_utils.FisheyeCameraCfg( clipping_range=(0.05, 3.0), projection_type="fisheyePolynomial", focal_length=23.151, horizontal_aperture=100.0, vertical_aperture=62.5, fisheye_nominal_width=960.0, fisheye_nominal_height=600.0, fisheye_optical_centre_x=483.472, fisheye_optical_centre_y=299.437, fisheye_max_fov=180.0, fisheye_polynomial_a=0.0, fisheye_polynomial_b=0.00245, fisheye_polynomial_c=0.0, fisheye_polynomial_d=0.0, fisheye_polynomial_e=0.0, fisheye_polynomial_f=0.0, ), height=600, width=960, ) # Right arm, minus-side (BFS_23595722) self.scene.flir_wrist_right_minus = TiledCameraCfg( prim_path="{ENV_REGEX_NS}/right_panda/panda_link8/flir_wrist_right_minus", offset=TiledCameraCfg.OffsetCfg( pos=(-0.04651, 0.04651, 0.0048), rot=(0.38268, 0.0, 0.0, -0.92388), convention="ros", ), data_types=["rgb"], spawn=sim_utils.FisheyeCameraCfg( clipping_range=(0.05, 3.0), projection_type="fisheyePolynomial", focal_length=18.627, horizontal_aperture=100.0, vertical_aperture=62.5, fisheye_nominal_width=960.0, fisheye_nominal_height=600.0, fisheye_optical_centre_x=459.918, fisheye_optical_centre_y=301.923, fisheye_max_fov=180.0, fisheye_polynomial_a=0.0, fisheye_polynomial_b=0.00245, fisheye_polynomial_c=0.0, fisheye_polynomial_d=0.0, fisheye_polynomial_e=0.0, fisheye_polynomial_f=0.0, ), height=600, width=960, ) self.observations.vision = VisionCfg() @configclass class ScenarioRelativeIKStateCfg(ScenarioIKStateCfg): def __post_init__(self): super().__post_init__() self.actions.dual_arm_ik_action.controller.use_relative_mode = True self.scene.left_panda.spawn.rigid_props.disable_gravity = True self.scene.right_panda.spawn.rigid_props.disable_gravity = True @configclass class ScenarioRelativeIKVisionCfg(ScenarioIkVisionCfg): def __post_init__(self): super().__post_init__() self.actions.dual_arm_ik_action.controller.use_relative_mode = True self.scene.left_panda.spawn.rigid_props.disable_gravity = True self.scene.right_panda.spawn.rigid_props.disable_gravity = True # del self.scene.flir_wrist_left_plus # del self.scene.flir_wrist_right_minus # del self.observations.vision