ignore collisions

Author: beneisner

notebooks/explore_dset2.ipynb

@@ -18,7 +18,7 @@
    "source": [
     "from rpad.rlbench_utils.placement_dataset import RLBenchPlacementDataset, load_handle_mapping, load_state_pos_dict, TASK_DICT\n",
     "import numpy as np\n",
-    "\n",
+    "from rpad.rlbench_utils.task_info import RLBENCH_10_TASKS\n",
     "from rpad.visualize_3d.plots import segmentation_fig\n",
     "import torch\n",
     "import matplotlib.pyplot as plt\n",
@@ -32,7 +32,7 @@
    "outputs": [],
    "source": [
     "dset = RLBenchPlacementDataset(\n",
-    "    dataset_root=\"/data/rlbench10/\",\n",
+    "    dataset_root=\"/data/rlbench10_collisions/\",\n",
     "    # task_name=\"stack_wine\",\n",
     "    # task_name=\"insert_onto_square_peg\",\n",
     "    # task_name=\"insert_usb_in_computer\",\n",
@@ -42,11 +42,38 @@
     "    # task_name=\"solve_puzzle\",\n",
     "    # task_name=\"take_umbrella_out_of_umbrella_stand\",\n",
     "    task_name=\"slide_block_to_target\",\n",
-    "    demos=range(10),\n",
+    "    demos=range(100),\n",
     "    phase=\"all\",\n",
     ")"
    ]
   },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "data = dset[0]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "data[\"ignore_collisions\"]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "data.keys()"
+   ]
+  },
   {
    "cell_type": "code",
    "execution_count": null,
@@ -707,6 +734,45 @@
     "unique_elements"
    ]
   },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "for task_name in RLBENCH_10_TASKS:\n",
+    "    print(\"--------------------\")\n",
+    "    print(f\"Task: {task_name}\")\n",
+    "    print(\"--------------------\")\n",
+    "    for phase in TASK_DICT[task_name][\"phase\"].keys():\n",
+    "\n",
+    "        dset = RLBenchPlacementDataset(\n",
+    "            dataset_root=\"/data/rlbench10_collisions/\",\n",
+    "            # task_name=\"stack_wine\",\n",
+    "            # task_name=\"insert_onto_square_peg\",\n",
+    "            # task_name=\"insert_usb_in_computer\",\n",
+    "            # task_name=\"phone_on_base\",\n",
+    "            # task_name=\"put_toilet_roll_on_stand\",\n",
+    "            # task_name=\"place_hanger_on_rack\",\n",
+    "            # task_name=\"solve_puzzle\",\n",
+    "            # task_name=\"take_umbrella_out_of_umbrella_stand\",\n",
+    "            task_name=task_name,\n",
+    "            demos=range(100),\n",
+    "            phase=phase,\n",
+    "        )   \n",
+    "        ignore_collisions_all = []\n",
+    "        for i in range(len(dset)):\n",
+    "            try:\n",
+    "                data = dset[i]\n",
+    "                ignore_collisions_all.append(data[\"ignore_collisions\"])\n",
+    "            except:\n",
+    "                print(f\"Error in task {task_name}, phase {phase}, demo {i}\")\n",
+    "        ignore_all = (np.array(ignore_collisions_all).any())    \n",
+    "        print(f\"Phase: {phase}; Ignore Collisions: {ignore_all}\")\n",
+    "\n",
+    "\n"
+   ]
+  },
   {
    "cell_type": "code",
    "execution_count": null,

src/rpad/rlbench_utils/placement_dataset.py

@@ -467,6 +467,12 @@ def extract_pose(obs, key):
         T_init_key = T_action_key_world @ np.linalg.inv(T_action_init_world)
         T_anchor_key_world = extract_pose(key_obs, "anchor_pose_name")
 
+        if hasattr(initial_obs, "ignore_collisions"):
+            ignore_collisions = initial_obs.ignore_collisions
+            ignore_collisions = torch.from_numpy(ignore_collisions.astype(np.int32))
+        else:
+            ignore_collisions = None
+
         return {
             "init_action_rgb": torch.from_numpy(init_action_rgb),
             "init_action_pc": torch.from_numpy(init_action_point_cloud),
@@ -489,4 +495,5 @@ def extract_pose(obs, key):
             "key_front_mask": torch.from_numpy(key_obs.front_mask.astype(np.int32)),
             "phase": phase,
             "phase_onehot": torch.from_numpy(phase_onehot),
+            "ignore_collisions": ignore_collisions,
         }