Episode Creator Script for Habitat 2.0 (facebookresearch#764)

* Introduce RearrangeEpisodeGenerator with redesigned RearrangeEpisode and RearrangeDataset for generating small object clutter within Receptacles fixed to objects or defined in the global frame. * Various Habitat-sim utilities supporting episode generation and debugging. * Refactored rearrange tasks to fix bugs, refine implementation and support new episode format. * Gym API for rearrange task training. * Updated rearrange task and episode generation config files. * Improved interactive_play application for interactive task design and debugging. * Refactored unit tests to better cover rearrange episode generation and task execution. Co-authored-by: Alexander William Clegg <alexclegg@fb.com> Co-authored-by: Oleksandr Maksymets <maksymets@gmail.com> Co-authored-by: Naoki Yokoyama <naokiyokoyama@devfair0196.h2.fair>
SgtVincent · Dec 10, 2021 · e7d746a · e7d746a
1 parent d6a1930
commit e7d746a
Show file tree

Hide file tree

Showing 59 changed files with 4,869 additions and 5,720 deletions.
diff --git a/README.md b/README.md
@@ -107,14 +107,14 @@ If you use the Habitat platform in your research, please cite the [Habitat](http
 ## Example
 <!--- Please, update `examples/example.py` if you update example. -->
 
-🆕Example code-snippet which uses [`tasks/rearrangepick_replica_cad.yaml`](configs/tasks/rearrangepick_replica_cad.yaml) for configuration of task and agent.
+🆕Example code-snippet which uses [`configs/tasks/rearrange/pick.yaml`](configs/tasks/rearrange/pick.yaml) for configuration of task and agent.
 
 ```python
 import habitat
 
 # Load embodied AI task (RearrangePick) and a pre-specified virtual robot
 env = habitat.Env(
-    config=habitat.get_config("configs/tasks/rearrangepick_replica_cad.yaml")
+    config=habitat.get_config("configs/tasks/rearrange/pick.yaml")
 )
 
 observations = env.reset()

diff --git a/configs/tasks/rearrangepick_replica_cad.yaml → configs/tasks/rearrange/pick.yaml b/configs/tasks/rearrangepick_replica_cad.yaml → configs/tasks/rearrange/pick.yaml
@@ -3,17 +3,24 @@ ENVIRONMENT:
 DATASET:
     TYPE: RearrangeDataset-v0
     SPLIT: train
-    DATA_PATH: data/datasets/rearrange_pick/replica_cad/v0/{split}/{split}_counter_L_analysis_5000_500.json.gz
+    DATA_PATH: data/datasets/rearrange_pick/replica_cad/v0/rearrange_pick_replica_cad_v0/pick.json.gz
     SCENES_DIR: "data/replica_cad/"
 TASK:
     TYPE: RearrangePickTask-v0
     MAX_COLLISIONS: -1.0
     COUNT_OBJ_COLLISIONS: True
     COUNT_ROBOT_OBJ_COLLS: False
+    DESIRED_RESTING_POSITION: [0.5, 0.0, 1.0]
 
     # In radians
     BASE_ANGLE_NOISE: 0.15
     BASE_NOISE: 0.05
+    CONSTRAINT_VIOLATION_ENDS_EPISODE: True
+    FORCE_REGENERATE: False
+
+    # Measurements for composite tasks.
+    REWARD_MEASUREMENT: "rearrangepick_reward"
+    SUCCESS_MEASUREMENT: "rearrangepick_success"
 
     # If true, does not care about navigability or collisions with objects when spawning
     # robot
@@ -44,52 +51,52 @@ TASK:
         TYPE: "IsHoldingSensor"
     RELATIVE_RESTING_POS_SENSOR:
         TYPE: "RelativeRestingPositionSensor"
-    LOCAL_SENSOR:
-        TYPE: "LocalizationSensor"
-    SENSORS: ["TARGET_START_SENSOR", "GOAL_SENSOR", "JOINT_SENSOR", "IS_HOLDING_SENSOR", "END_EFFECTOR_SENSOR", "LOCAL_SENSOR", "RELATIVE_RESTING_POS_SENSOR"]
+    SENSORS: ["TARGET_START_SENSOR", "JOINT_SENSOR", "IS_HOLDING_SENSOR", "END_EFFECTOR_SENSOR", "RELATIVE_RESTING_POS_SENSOR"]
     ROBOT_FORCE:
         TYPE: "RobotForce"
         MIN_FORCE: 20.0
+    EXCESSIVE_FORCE_SHOULD_END:
+        TYPE: "ForceTerminate"
+        MAX_ACCUM_FORCE: 5000.0
     ROBOT_COLLS:
-      TYPE: "RobotCollisions"
+        TYPE: "RobotCollisions"
     OBJECT_TO_GOAL_DISTANCE:
         TYPE: "ObjectToGoalDistance"
     END_EFFECTOR_TO_OBJECT_DISTANCE:
         TYPE: "EndEffectorToObjectDistance"
     END_EFFECTOR_TO_REST_DISTANCE:
         TYPE: "EndEffectorToRestDistance"
-    REARRANGE_PICK_REWARD:
+    PICK_REWARD:
         TYPE: "RearrangePickReward"
         DIST_REWARD: 20.0
         SUCC_REWARD: 10.0
         PICK_REWARD: 20.0
-        CONSTRAINT_VIOLATE_PEN: 10.0
         DROP_PEN: 5.0
         WRONG_PICK_PEN: 5.0
-        COLL_PEN: 1.0
-        ROBOT_OBJ_COLL_PEN: 0.0
-        MAX_ACCUM_FORCE: 5000.0
+        USE_DIFF: True
+        DROP_OBJ_SHOULD_END: False
+        WRONG_PICK_SHOULD_END: False
 
+        # General Rearrange Reward config
+        CONSTRAINT_VIOLATE_PEN: 10.0
         FORCE_PEN: 0.001
         MAX_FORCE_PEN: 1.0
         FORCE_END_PEN: 10.0
-        USE_DIFF: True
 
-        DROP_OBJ_SHOULD_END: False
-        WRONG_PICK_SHOULD_END: False
 
-    REARRANGE_PICK_SUCCESS:
+    PICK_SUCCESS:
         TYPE: "RearrangePickSuccess"
         SUCC_THRESH: 0.15
 
     MEASUREMENTS:
         - "OBJECT_TO_GOAL_DISTANCE"
         - "ROBOT_FORCE"
+        - "EXCESSIVE_FORCE_SHOULD_END"
         - "ROBOT_COLLS"
         - "END_EFFECTOR_TO_REST_DISTANCE"
         - "END_EFFECTOR_TO_OBJECT_DISTANCE"
-        - "REARRANGE_PICK_SUCCESS"
-        - "REARRANGE_PICK_REWARD"
+        - "PICK_SUCCESS"
+        - "PICK_REWARD"
     ACTIONS:
         ARM_ACTION:
             TYPE: "ArmAction"
@@ -99,12 +106,14 @@ TASK:
             GRASP_THRESH_DIST: 0.15
             DISABLE_GRIP: False
             DELTA_POS_LIMIT: 0.0125
+            EE_CTRL_LIM: 0.015
     POSSIBLE_ACTIONS:
         - ARM_ACTION
 
 SIMULATOR:
     ACTION_SPACE_CONFIG: v0
     AGENTS: ['AGENT_0']
+    ROBOT_JOINT_START_NOISE: 0.0
     AGENT_0:
         HEIGHT: 1.5
         IS_SET_START_STATE: False
@@ -136,8 +145,8 @@ SIMULATOR:
     CTRL_FREQ: 120.0
     AC_FREQ_RATIO: 4
     ROBOT_URDF: ./data/robots/hab_fetch/robots/hab_fetch.urdf
+    ROBOT_TYPE: "FetchRobot"
     FORWARD_STEP_SIZE: 0.25
-    LAG_OBSERVATIONS: 1
 
     # Grasping
     HOLD_THRESH: 0.09

diff --git a/.../tasks/rearrangepick_replica_cad_spa.yaml → configs/tasks/rearrange/pick_spa.yaml b/.../tasks/rearrangepick_replica_cad_spa.yaml → configs/tasks/rearrange/pick_spa.yaml
@@ -1,21 +1,23 @@
 # The configuration to run the SensePlanAct baseline architecture from https://arxiv.org/abs/2106.14405
 ENVIRONMENT:
-    MAX_EPISODE_STEPS: 200
+    MAX_EPISODE_STEPS: 30
 DATASET:
     TYPE: RearrangeDataset-v0
     SPLIT: train
-    DATA_PATH: data/datasets/rearrange_pick/replica_cad/v0/{split}/{split}_counter_L_analysis_5000_500.json.gz
+    DATA_PATH: data/datasets/rearrange_pick/replica_cad/v0/rearrange_pick_replica_cad_v0/pick.json.gz
     SCENES_DIR: "data/replica_cad/"
 TASK:
     TYPE: RearrangePickTask-v0
     MAX_COLLISIONS: -1.0
     COUNT_OBJ_COLLISIONS: True
     COUNT_ROBOT_OBJ_COLLS: False
+    CONSTRAINT_VIOLATION_ENDS_EPISODE: True
 
     # In radians
     BASE_ANGLE_NOISE: 0.15
     BASE_NOISE: 0.05
-    FORCE_RECACHE: False
+    FORCE_REGENERATE: False
+    DESIRED_RESTING_POSITION: [0.5, 0.0, 1.0]
 
     # If true, does not care about navigability or collisions with objects when spawning
     # robot
@@ -48,11 +50,9 @@ TASK:
         TYPE: "EEPositionSensor"
     IS_HOLDING_SENSOR:
         TYPE: "IsHoldingSensor"
-    LOCAL_SENSOR:
-        TYPE: "LocalizationSensor"
     RELATIVE_RESTING_POS_SENSOR:
         TYPE: "RelativeRestingPositionSensor"
-    SENSORS: ["TARGET_START_SENSOR", "GOAL_SENSOR", "JOINT_SENSOR", "IS_HOLDING_SENSOR", "END_EFFECTOR_SENSOR", "LOCAL_SENSOR", "TARGET_CUR_SENSOR", "TARGET_CUR_SENSOR", "RELATIVE_RESTING_POS_SENSOR"]
+    SENSORS: ["TARGET_START_SENSOR", "GOAL_SENSOR", "JOINT_SENSOR", "IS_HOLDING_SENSOR", "END_EFFECTOR_SENSOR", "TARGET_CUR_SENSOR", "TARGET_CUR_SENSOR", "RELATIVE_RESTING_POS_SENSOR"]
     ROBOT_FORCE:
         TYPE: "RobotForce"
         MIN_FORCE: 20.0
@@ -64,37 +64,37 @@ TASK:
         TYPE: "EndEffectorToObjectDistance"
     END_EFFECTOR_TO_REST_DISTANCE:
         TYPE: "EndEffectorToRestDistance"
-    REARRANGE_PICK_REWARD:
+    EXCESSIVE_FORCE_SHOULD_END:
+        TYPE: "ForceTerminate"
+        MAX_ACCUM_FORCE: 0.0
+    PICK_REWARD:
         TYPE: "RearrangePickReward"
-        DIST_REWARD: 20.0
-        SUCC_REWARD: 10.0
-        PICK_REWARD: 20.0
+        DIST_REWARD: 10.0
+        PICK_REWARD: 5.0
         CONSTRAINT_VIOLATE_PEN: 10.0
         DROP_PEN: 5.0
         WRONG_PICK_PEN: 5.0
-        COLL_PEN: 1.0
-        ROBOT_OBJ_COLL_PEN: 0.0
-        MAX_ACCUM_FORCE: null
+        COLLISION_PENALTY: 0.0
+        ROBOT_OBJ_COLLISION_PENALTY: 0.0
 
-        FORCE_PEN: 0.001
-        MAX_FORCE_PEN: 1.0
-        FORCE_END_PEN: 10.0
+        FORCE_PEN: 0.0
+        MAX_FORCE_PEN: 0.0
+        FORCE_END_PEN: 0.0
         USE_DIFF: True
 
         DROP_OBJ_SHOULD_END: False
         WRONG_PICK_SHOULD_END: False
-
-    REARRANGE_PICK_SUCCESS:
+    PICK_SUCCESS:
         TYPE: "RearrangePickSuccess"
-        SUCC_THRESH: 0.3
-
+        SUCC_THRESH: 0.15
     MEASUREMENTS:
         - "ROBOT_FORCE"
         - "ROBOT_COLLS"
-        - "END_EFFECTOR_TO_OBJECT_DISTANCE"
+        - "EXCESSIVE_FORCE_SHOULD_END"
         - "END_EFFECTOR_TO_REST_DISTANCE"
-        - "REARRANGE_PICK_SUCCESS"
-        - "REARRANGE_PICK_REWARD"
+        - "END_EFFECTOR_TO_OBJECT_DISTANCE"
+        - "PICK_SUCCESS"
+        - "PICK_REWARD"
     ACTIONS:
         ARM_ACTION:
             TYPE: "ArmAction"
@@ -105,6 +105,7 @@ TASK:
             DISABLE_GRIP: False
             DELTA_POS_LIMIT: 0.0125
             EE_CTRL_LIM: 0.015
+            ORACLE_GRASP: True
         EMPTY:
             TYPE: "EmptyAction"
     POSSIBLE_ACTIONS:
@@ -114,6 +115,7 @@ TASK:
 SIMULATOR:
     ACTION_SPACE_CONFIG: v0
     AGENTS: ['AGENT_0']
+    ROBOT_JOINT_START_NOISE: 0.0
     AGENT_0:
         HEIGHT: 1.5
         IS_SET_START_STATE: False
@@ -146,11 +148,11 @@ SIMULATOR:
     # Agent setup
     ARM_REST: [0.6, 0.0, 0.9]
     CTRL_FREQ: 120.0
-    AC_FREQ_RATIO: 4
+    AC_FREQ_RATIO: 8
     ROBOT_URDF: ./data/robots/hab_fetch/robots/hab_fetch.urdf
+    ROBOT_TYPE: "FetchRobot"
     IK_ARM_URDF: ./data/robots/hab_fetch/robots/fetch_onlyarm.urdf
     FORWARD_STEP_SIZE: 0.25
-    LAG_OBSERVATIONS: 1
 
     # Grasping
     HOLD_THRESH: 0.09