🚀 [RofuncRL] Add QbSofthand-based tasks

examples/learning_rl/IsaacGym_RofuncRL/example_DexterousHands_RofuncRL.py

@@ -97,12 +97,13 @@ def inference(custom_args):
     #                  BiShadowHandGraspAndPlace, BiShadowHandLiftUnderarm, BiShadowHandPen, BiShadowHandPointCloud,
     #                  BiShadowHandPushBlock, BiShadowHandReOrientation, BiShadowHandScissors, BiShadowHandSwingCup,
     #                  BiShadowHandSwitch, BiShadowHandTwoCatchUnderarm
-    parser.add_argument("--task", type=str, default="BiShadowHandSwitch")
+    #                  QbSoftHandGrasp
+    parser.add_argument("--task", type=str, default="BiQbSoftHandGraspAndPlace")
     parser.add_argument("--agent", type=str, default="ppo")  # Available agents: ppo, sac, td3, a2c
     parser.add_argument("--num_envs", type=int, default=256)
     parser.add_argument("--sim_device", type=int, default=0)
     parser.add_argument("--rl_device", type=int, default=gpu_id)
-    parser.add_argument("--headless", type=str, default="True")
+    parser.add_argument("--headless", type=str, default="False")
     parser.add_argument("--inference", action="store_true", help="turn to inference mode while adding this argument")
     parser.add_argument("--ckpt_path", type=str, default=None)
     custom_args = parser.parse_args()

rofunc/config/learning/rl/task/BiQbSoftHandGraspAndPlace.yaml

@@ -0,0 +1,178 @@
+name: BiShadowHandGraspAndPlace
+
+physics_engine: ${..physics_engine}
+
+# if given, will override the device setting in gym.
+env: 
+  env_name: "bi_qbsofthand_grasp_and_place"
+  numEnvs: ${resolve_default:4096,${...num_envs}}
+  envSpacing: 1.5
+  episodeLength: 500
+  enableDebugVis: False
+  cameraDebug: True
+  pointCloudDebug: True
+  aggregateMode: 1
+
+  stiffnessScale: 1.0
+  forceLimitScale: 1.0
+  useRelativeControl: False
+  dofSpeedScale: 20.0
+  actionsMovingAverage: 1.0
+  controlFrequencyInv: 1 # 60 Hz
+
+  startPositionNoise: 0.0
+  startRotationNoise: 0.0
+
+  resetPositionNoise: 0.0
+  resetRotationNoise: 0.0
+  resetDofPosRandomInterval: 0.0
+  resetDofVelRandomInterval: 0.0
+
+  distRewardScale: 20
+  transition_scale: 0.5
+  orientation_scale: 0.1
+  rotRewardScale: 1.0
+  rotEps: 0.1
+  actionPenaltyScale: -0.0002
+  reachGoalBonus: 250
+  fallDistance: 0.4
+  fallPenalty: 0.0
+
+  objectType: "pot" # can be block, egg or pen
+  observationType: "full_state" # point_cloud or full_state
+  handAgentIndex: "[[0, 1, 2, 3, 4, 5]]"
+  asymmetric_observations: False
+  successTolerance: 0.1
+  printNumSuccesses: False
+  maxConsecutiveSuccesses: 0
+
+  asset:
+    assetFileNameBlock: "urdf/objects/cube_multicolor.urdf"
+    assetFileNameEgg: "mjcf/open_ai_assets/hand/egg.xml"
+    assetFileNamePen: "mjcf/open_ai_assets/hand/pen.xml"
+
+task:
+  randomize: False
+  randomization_params:
+    frequency: 600   # Define how many simulation steps between generating new randomizations
+    observations:
+      range: [0, .002] # range for the white noise
+      range_correlated: [0, .001 ] # range for correlated noise, refreshed with freq `frequency`
+      operation: "additive"
+      distribution: "gaussian"
+      schedule: "linear"   # "constant" is to turn on noise after `schedule_steps` num steps
+      schedule_steps: 40000
+    actions:
+      range: [0., .05]
+      range_correlated: [0, .015] # range for correlated noise, refreshed with freq `frequency`
+      operation: "additive"
+      distribution: "gaussian"
+      schedule: "linear"  # "linear" will linearly interpolate between no rand and max rand
+      schedule_steps: 40000
+    sim_params: 
+      gravity:
+        range: [0, 0.4]
+        operation: "additive"
+        distribution: "gaussian"
+        schedule: "linear"  # "linear" will linearly interpolate between no rand and max rand
+        schedule_steps: 40000
+    actor_params:
+      hand:
+        color: True
+        tendon_properties:
+          damping:
+            range: [0.3, 3.0]
+            operation: "scaling"
+            distribution: "loguniform"
+            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+          stiffness:
+            range: [0.75, 1.5]
+            operation: "scaling"
+            distribution: "loguniform"
+            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+        dof_properties:
+          damping: 
+            range: [0.3, 3.0]
+            operation: "scaling"
+            distribution: "loguniform"
+            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+          stiffness: 
+            range: [0.75, 1.5]
+            operation: "scaling"
+            distribution: "loguniform"
+            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+          lower:
+            range: [0, 0.01]
+            operation: "additive"
+            distribution: "gaussian"
+            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+          upper:
+            range: [0, 0.01]
+            operation: "additive"
+            distribution: "gaussian"
+            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+        rigid_body_properties:
+          mass: 
+            range: [0.5, 1.5]
+            operation: "scaling"
+            distribution: "uniform"
+            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+        rigid_shape_properties:
+          friction: 
+            num_buckets: 250
+            range: [0.7, 1.3]
+            operation: "scaling"
+            distribution: "uniform"
+            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+      object:
+        scale:
+          range: [0.95, 1.05]
+          operation: "scaling"
+          distribution: "uniform"
+          schedule: "linear"  # "linear" will scale the current random sample by ``min(current num steps, schedule_steps) / schedule_steps`
+          schedule_steps: 30000
+        rigid_body_properties:
+          mass: 
+            range: [0.5, 1.5]
+            operation: "scaling"
+            distribution: "uniform"
+            schedule: "linear"  # "linear" will scale the current random sample by ``min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+        rigid_shape_properties:
+          friction:
+            num_buckets: 250
+            range: [0.7, 1.3]
+            operation: "scaling"
+            distribution: "uniform"
+            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+
+sim:
+  dt: 0.0166 # 1/60 s
+  substeps: 2
+  up_axis: "z"
+  use_gpu_pipeline: ${eq:${...pipeline},"gpu"}
+  gravity: [ 0.0, 0.0, -9.81 ]
+  physx:
+    num_threads: 4
+    solver_type: 1  # 0: pgs, 1: tgs
+    num_position_iterations: 8
+    num_velocity_iterations: 0
+    contact_offset: 0.002
+    rest_offset: 0.0
+    bounce_threshold_velocity: 0.2
+    max_depenetration_velocity: 1000.0
+    default_buffer_size_multiplier: 5.0
+  flex:
+    num_outer_iterations: 5
+    num_inner_iterations: 20
+    warm_start: 0.8
+    relaxation: 0.75

rofunc/config/learning/rl/task/QbSoftHandGrasp.yaml

@@ -0,0 +1,179 @@
+name: QbSoftHandGrasp
+
+physics_engine: ${..physics_engine}
+
+# if given, will override the device setting in gym.
+env: 
+  env_name: "qbsofthand_grasp"
+  numEnvs: ${resolve_default:4096,${...num_envs}}
+  envSpacing: 1.5
+  episodeLength: 500
+  enableDebugVis: False
+  cameraDebug: True
+  pointCloudDebug: True
+  aggregateMode: 1
+
+  stiffnessScale: 1.0
+  forceLimitScale: 1.0
+  useRelativeControl: False
+  dofSpeedScale: 20.0
+  actionsMovingAverage: 1.0
+  controlFrequencyInv: 1 # 60 Hz
+
+  startPositionNoise: 0.0
+  startRotationNoise: 0.0
+
+  resetPositionNoise: 0.0
+  resetRotationNoise: 0.0
+  resetDofPosRandomInterval: 0.0
+  resetDofVelRandomInterval: 0.0
+
+  distRewardScale: 20
+  transition_scale: 0.5
+  orientation_scale: 0.1
+  rotRewardScale: 1.0
+  rotEps: 0.1
+  actionPenaltyScale: -0.0002
+  reachGoalBonus: 250
+  fallDistance: 0.4
+  fallPenalty: 0.0
+
+  objectType: "pot" # can be block, egg or pen
+  observationType: "full_state" # point_cloud or full_state
+  handAgentIndex: "[[0, 1, 2, 3, 4, 5]]"
+  asymmetric_observations: False
+  successTolerance: 0.1
+  printNumSuccesses: False
+  maxConsecutiveSuccesses: 0
+
+  asset:
+#    assetFileName: "mjcf/open_ai_assets/hand/shadow_hand.xml"
+    assetFileNameBlock: "urdf/objects/cube_multicolor.urdf"
+    assetFileNameEgg: "mjcf/open_ai_assets/hand/egg.xml"
+    assetFileNamePen: "mjcf/open_ai_assets/hand/pen.xml"
+
+task:
+  randomize: False
+  randomization_params:
+    frequency: 600   # Define how many simulation steps between generating new randomizations
+    observations:
+      range: [0, .002] # range for the white noise
+      range_correlated: [0, .001 ] # range for correlated noise, refreshed with freq `frequency`
+      operation: "additive"
+      distribution: "gaussian"
+      schedule: "linear"   # "constant" is to turn on noise after `schedule_steps` num steps
+      schedule_steps: 40000
+    actions:
+      range: [0., .05]
+      range_correlated: [0, .015] # range for correlated noise, refreshed with freq `frequency`
+      operation: "additive"
+      distribution: "gaussian"
+      schedule: "linear"  # "linear" will linearly interpolate between no rand and max rand
+      schedule_steps: 40000
+    sim_params: 
+      gravity:
+        range: [0, 0.4]
+        operation: "additive"
+        distribution: "gaussian"
+        schedule: "linear"  # "linear" will linearly interpolate between no rand and max rand
+        schedule_steps: 40000
+    actor_params:
+      hand:
+        color: True
+        tendon_properties:
+          damping:
+            range: [0.3, 3.0]
+            operation: "scaling"
+            distribution: "loguniform"
+            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+          stiffness:
+            range: [0.75, 1.5]
+            operation: "scaling"
+            distribution: "loguniform"
+            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+        dof_properties:
+          damping: 
+            range: [0.3, 3.0]
+            operation: "scaling"
+            distribution: "loguniform"
+            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+          stiffness: 
+            range: [0.75, 1.5]
+            operation: "scaling"
+            distribution: "loguniform"
+            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+          lower:
+            range: [0, 0.01]
+            operation: "additive"
+            distribution: "gaussian"
+            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+          upper:
+            range: [0, 0.01]
+            operation: "additive"
+            distribution: "gaussian"
+            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+        rigid_body_properties:
+          mass: 
+            range: [0.5, 1.5]
+            operation: "scaling"
+            distribution: "uniform"
+            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+        rigid_shape_properties:
+          friction: 
+            num_buckets: 250
+            range: [0.7, 1.3]
+            operation: "scaling"
+            distribution: "uniform"
+            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+      object:
+        scale:
+          range: [0.95, 1.05]
+          operation: "scaling"
+          distribution: "uniform"
+          schedule: "linear"  # "linear" will scale the current random sample by ``min(current num steps, schedule_steps) / schedule_steps`
+          schedule_steps: 30000
+        rigid_body_properties:
+          mass: 
+            range: [0.5, 1.5]
+            operation: "scaling"
+            distribution: "uniform"
+            schedule: "linear"  # "linear" will scale the current random sample by ``min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+        rigid_shape_properties:
+          friction:
+            num_buckets: 250
+            range: [0.7, 1.3]
+            operation: "scaling"
+            distribution: "uniform"
+            schedule: "linear"  # "linear" will scale the current random sample by `min(current num steps, schedule_steps) / schedule_steps`
+            schedule_steps: 30000
+
+sim:
+  dt: 0.0166 # 1/60 s
+  substeps: 2
+  up_axis: "z"
+  use_gpu_pipeline: ${eq:${...pipeline},"gpu"}
+  gravity: [ 0.0, 0.0, -9.81 ]
+  physx:
+    num_threads: 4
+    solver_type: 1  # 0: pgs, 1: tgs
+    num_position_iterations: 8
+    num_velocity_iterations: 0
+    contact_offset: 0.002
+    rest_offset: 0.0
+    bounce_threshold_velocity: 0.2
+    max_depenetration_velocity: 1000.0
+    default_buffer_size_multiplier: 5.0
+  flex:
+    num_outer_iterations: 5
+    num_inner_iterations: 20
+    warm_start: 0.8
+    relaxation: 0.75

rofunc/learning/RofuncRL/tasks/__init__.py

@@ -40,6 +40,8 @@ def __init__(self, env_type="isaacgym"):
             from .isaacgymenv.hands.shadow_hand_swing_cup import ShadowHandSwingCupTask
             from .isaacgymenv.hands.shadow_hand_switch import ShadowHandSwitchTask
             from .isaacgymenv.hands.shadow_hand_two_catch_underarm import ShadowHandTwoCatchUnderarmTask
+            from .isaacgymenv.hands.qbsofthand_grasp import QbSoftHandGraspTask
+            from .isaacgymenv.hands.bi_qbhand_grasp_and_place import BiQbSoftHandGraspAndPlaceTask
 
             self.task_map = {
                 "Ant": AntTask,
@@ -80,6 +82,8 @@ def __init__(self, env_type="isaacgym"):
                 "BiShadowHandSwingCup": ShadowHandSwingCupTask,
                 "BiShadowHandSwitch": ShadowHandSwitchTask,
                 "BiShadowHandTwoCatchUnderarm": ShadowHandTwoCatchUnderarmTask,
+                "QbSoftHandGrasp": QbSoftHandGraspTask,
+                "BiQbSoftHandGraspAndPlace": BiQbSoftHandGraspAndPlaceTask,
             }
         elif env_type == "omniisaacgym":
             # OmniIsaacGym tasks