🚀 [RofuncRL] CURI Synergy-based Softhand grasping tasks are supported

changelog.md

@@ -11,6 +11,7 @@
 > - 🎉 for event celebrations
 
 ## Update News 🎉🎉🎉
+- [2024-01-24] 🚀 [CURI Synergy-based Softhand grasping tasks](https://github.com/Skylark0924/Rofunc/blob/main/examples/learning_rl/IsaacGym_RofuncRL/example_DexterousHands_RofuncRL.py) are supported to be trained by `RofuncRL`.
 - [2023-12-24] 🚀 [Dexterous hand (Shadow Hand, Allegro Hand, qbSofthand) tasks](https://github.com/Skylark0924/Rofunc/blob/main/examples/learning_rl/IsaacGym_RofuncRL/example_DexterousHands_RofuncRL.py) are supported to be trained by `RofuncRL`.
 - [2023-12-07] 🖼️ [EfficientSAM](https://yformer.github.io/efficient-sam/) is supported for high-speed segmentation on edge devices like Nvidia Jetson, check the [example](https://github.com/Skylark0924/Rofunc/blob/main/examples/visualab/example_efficient_sam_seg_w_prompt.py) in Visualab.
 - [2023-12-04] 🖼️ [VLPart-SAM](https://github.com/Cheems-Seminar/grounded-segment-any-parts) is supported for part-level segmentation with text prompt, check the [example](https://github.com/Skylark0924/Rofunc/blob/main/examples/visualab/example_vlpart_sam_seg_w_prompt.py).

examples/learning_rl/IsaacGym_RofuncRL/example_DexterousHands_RofuncRL.py

@@ -7,6 +7,7 @@
 
 import isaacgym
 import argparse
+import datetime
 
 from rofunc.config.utils import omegaconf_to_dict, get_config
 from rofunc.learning.RofuncRL.tasks import Tasks
@@ -26,9 +27,10 @@ def train(custom_args):
                       "headless={}".format(custom_args.headless),
                       "num_envs={}".format(custom_args.num_envs)]
     cfg = get_config('./learning/rl', 'config', args=args_overrides)
-    cfg.task.env.objectType = custom_args.object.lower()
+    if custom_args.task == "CURIQbSoftHandSynergyGrasp":
+        cfg.task.env.objectType = custom_args.object.lower()
     cfg.train.Trainer.maximum_steps = 100000
-    cfg.train.Trainer.experiment_name = custom_args.object
+    cfg.train.Trainer.experiment_name = "{}_{}".format(custom_args.object, datetime.datetime.now().strftime("%y-%m-%d_%H-%M-%S-%f"))
     cfg_dict = omegaconf_to_dict(cfg.task)
 
     set_seed(cfg.train.Trainer.seed)
@@ -47,6 +49,8 @@ def train(custom_args):
                                                         env=env,
                                                         device=cfg.rl_device,
                                                         env_name=custom_args.task)
+    if custom_args.ckpt_path is not None:
+        trainer.agent.load_ckpt(custom_args.ckpt_path)
 
     # Start training
     trainer.train()
@@ -59,8 +63,13 @@ def inference(custom_args):
                       "device_id={}".format(custom_args.sim_device),
                       "rl_device=cuda:{}".format(custom_args.rl_device),
                       "headless={}".format(False),
-                      "num_envs={}".format(16)]
+                      "num_envs={}".format(1)]
     cfg = get_config('./learning/rl', 'config', args=args_overrides)
+    if custom_args.task == "CURIQbSoftHandSynergyGrasp":
+        cfg.task.env.objectType = custom_args.object.lower()
+    cfg.train.Trainer.experiment_name = custom_args.object
+    cfg.train.Trainer.inference_steps = 10000
+
     cfg_dict = omegaconf_to_dict(cfg.task)
 
     set_seed(cfg.train.Trainer.seed)
@@ -102,13 +111,17 @@ def inference(custom_args):
     #                  QbSoftHandGrasp, BiQbSoftHandGraspAndPlace, BiQbSoftHandSynergyGrasp, QbSoftHandSynergyGrasp
     #                  ShadowHandGrasp, CURIQbSoftHandSynergyGrasp
     parser.add_argument("--task", type=str, default="CURIQbSoftHandSynergyGrasp")
-    parser.add_argument("--object", type=str, default="Hammer")
+    # Available objects: Hammer, Spatula, Large_Clamp, Mug, Power_Drill, Knife, Scissors, Large_Marker, Phillips_Screw_Driver
+    # Only for CURIQbSoftHandSynergyGrasp
+    parser.add_argument("--object", type=str, default="Mug")
+    # Only for CURIQbSoftHandSynergyGrasp
+    parser.add_argument("--use_synergy", type=str, default="True")
     parser.add_argument("--agent", type=str, default="ppo")  # Available agents: ppo, sac, td3, a2c
     parser.add_argument("--num_envs", type=int, default=1024)
     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="False")
-    parser.add_argument("--inference", action="store_false", help="turn to inference mode while adding this argument")
+    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()
 

examples/learning_rl/README.rst

@@ -169,6 +169,10 @@ The following are examples of reinforcement learning methods for robot learning.
              - 
              - 
              -
+           * - `CURIQbSoftHandSynergyGrasp <https://github.com/Skylark0924/Rofunc/blob/main/rofunc/learning/RofuncRL/tasks/isaacgymenv/grasp/curi_qbhand_synergy_grasp.py>`_
+             - .. image:: ../../../img/CURIQbSoftHandSynergyGraspHammer.gif .. image:: ../../../img/CURIQbSoftHandSynergyGraspSpatula.gif .. image:: ../../../img/CURIQbSoftHandSynergyGraspPower_drill.gif .. image:: ../../../img/CURIQbSoftHandSynergyGraspPhillips_Screw_Driver.gif .. image:: ../../../img/CURIQbSoftHandSynergyGraspLarge_clamp.gif .. image:: ../../../img/CURIQbSoftHandSynergyGraspKnife.gif
+             -
+             - ✅
            * - `Humanoid <https://github.com/Skylark0924/Rofunc/blob/main/rofunc/learning/RofuncRL/tasks/isaacgymenv/humanoid.py>`_
              - .. image:: ../../../img/HumanoidRofuncRLPPO.gif
              -
@@ -299,11 +303,6 @@ The following are examples of reinforcement learning methods for robot learning.
              - ✅
 
 
-
-
-
-
-
     .. tab:: OmniIsaacGym
 
         .. tabs::

rofunc/config/learning/model_zoo.json

@@ -184,6 +184,61 @@
     "name": "CURIQbSoftHandSynergyGraspHammerRofuncRLPPO.pth",
     "md5": ""
   },
+  "_CURIQbSoftHandSynergyGraspLarge_Clamp2RofuncRLPPO.pth": {
+    "url": "https://drive.google.com/uc?id=1--hfNU1_o5x6tl2nb-lRJEzsEkcFD6_k&export=download",
+    "name": "CURIQbSoftHandSynergyGraspLarge_Clamp2RofuncRLPPO.pth",
+    "md5": ""
+  },
+  "CURIQbSoftHandSynergyGraspLarge_ClampRofuncRLPPO.pth": {
+    "url": "https://drive.google.com/uc?id=1Se-ghKuTfYgPFK1rNXhyaNqejkCuwlZ6&export=download",
+    "name": "CURIQbSoftHandSynergyGraspLarge_ClampRofuncRLPPO.pth",
+    "md5": ""
+  },
+  "CURIQbSoftHandSynergyGraspPhillips_Screw_DriverRofuncRLPPO.pth": {
+    "url": "https://drive.google.com/uc?id=1eM5m3qbhXIc9fZgI2AcFepwlcg7PAUEx&export=download",
+    "name": "CURIQbSoftHandSynergyGraspPhillips_Screw_DriverRofuncRLPPO.pth",
+    "md5": ""
+  },
+  "CURIQbSoftHandSynergyGraspLarge_MarkerRofuncRLPPO.pth": {
+    "url": "https://drive.google.com/uc?id=1EeTB-8wQTizjB-17YUdpa1dpH6uw0cyQ&export=download",
+    "name": "CURIQbSoftHandSynergyGraspLarge_MarkerRofuncRLPPO.pth",
+    "md5": ""
+  },
+  "_CURIQbSoftHandSynergyGraspMug2RofuncRLPPO.pth": {
+    "url": "https://drive.google.com/uc?id=1wI866QoCH8UTnh2X9SDeRh3pDH2R85-X&export=download",
+    "name": "CURIQbSoftHandSynergyGraspMug2RofuncRLPPO.pth",
+    "md5": ""
+  },
+  "_CURIQbSoftHandSynergyGraspMug3RofuncRLPPO.pth": {
+    "url": "https://drive.google.com/uc?id=13KkV_d9CVNzFMksswRivVgvKf4DY3f4i&export=download",
+    "name": "CURIQbSoftHandSynergyGraspMug3RofuncRLPPO.pth",
+    "md5": ""
+  },
+  "CURIQbSoftHandSynergyGraspMugRofuncRLPPO.pth": {
+    "url": "",
+    "name": "CURIQbSoftHandSynergyGraspMugRofuncRLPPO.pth",
+    "md5": ""
+  },
+  "CURIQbSoftHandSynergyGraspPower_DrillRofuncRLPPO.pth": {
+    "url": "https://drive.google.com/uc?id=1mxtw5GVEX_hDeGFTuXodeb5xBG2J5fBB&export=download",
+    "name": "CURIQbSoftHandSynergyGraspPower_DrillRofuncRLPPO.pth",
+    "md5": ""
+  },
+  "CURIQbSoftHandSynergyGraspKnifeRofuncRLPPO.pth": {
+    "url": "https://drive.google.com/uc?id=1Nihlp3PaJtUzhQMsDl871z_0imhurY4M&export=download",
+    "name": "CURIQbSoftHandSynergyGraspKnifeRofuncRLPPO.pth",
+    "md5": ""
+  },
+  "CURIQbSoftHandSynergyGraspSpatulaRofuncRLPPO.pth": {
+    "url": "https://drive.google.com/uc?id=15HdnTp_etpZFtp_b45cYoUFl8XLzsIbm&export=download",
+    "name": "CURIQbSoftHandSynergyGraspSpatulaRofuncRLPPO.pth",
+    "md5": ""
+  },
+  "CURIQbSoftHandSynergyGraspScissorsRofuncRLPPO.pth": {
+    "url": "https://drive.google.com/uc?id=1EeTB-8wQTizjB-17YUdpa1dpH6uw0cyQ&export=download",
+    "name": "CURIQbSoftHandSynergyGraspScissorsRofuncRLPPO.pth",
+    "md5": ""
+  },
   "efficientsam_s_gpu.jit": {
     "url": "https://drive.google.com/uc?id=1lMCUFOLeQOlIpuaIA00MB174lL7ZiuoS&export=download",
     "name": "efficientsam_s_gpu.jit",

rofunc/learning/RofuncRL/tasks/__init__.py

@@ -45,8 +45,7 @@ def __init__(self, env_type="isaacgym"):
             from .isaacgymenv.hands.bi_qbhand_synergy_grasp import BiQbSoftHandSynergyGraspTask
             from .isaacgymenv.hands.qbhand_synergy_grasp import QbSoftHandSynergyGraspTask
             from .isaacgymenv.hands.shadow_hand_grasp import ShadowHandGraspTask
-            from .isaacgymenv.grasp.lift_object import LiftObjectTask
-            from .isaacgymenv.hands.curi_qbhand_synergy_grasp import CURIQbSoftHandSynergyGraspTask
+            from .isaacgymenv.grasp.curi_qbhand_synergy_grasp import CURIQbSoftHandSynergyGraspTask
 
             self.task_map = {
                 "Ant": AntTask,
@@ -92,7 +91,6 @@ def __init__(self, env_type="isaacgym"):
                 "BiQbSoftHandSynergyGrasp": BiQbSoftHandSynergyGraspTask,
                 "QbSoftHandSynergyGrasp": QbSoftHandSynergyGraspTask,
                 "ShadowHandGrasp": ShadowHandGraspTask,
-                "LiftObject": LiftObjectTask,
                 "CURIQbSoftHandSynergyGrasp": CURIQbSoftHandSynergyGraspTask,
             }
         elif env_type == "omniisaacgym":