Add scenario tidy_stationery_gim

tidy_stationery_gim/media/clips.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5fe7e06b15c3db7fdac340d5f45cda91f31a9bedd845e813e7ab9df5a6d8f113
+size 811479040

tidy_stationery_gim/media/episodes.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e333e2fff0c91155cc6792493e528cd43fe9dc3539d1b5f24cd49a3a75c4136
+size 1033594880

tidy_stationery_gim/media/frames.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:119ed312e1eed9a995bfb2145f3d6e4e5af6e441afe58ccfc5366cd6a0b977ec
+size 75653120

tidy_stationery_gim/qa/dimensions.json

@@ -0,0 +1,13 @@
+{
+ "version": 1,
+ "_note": "Question templates, one per dimension. Slots: {subtask} the action this question is about, with its trailing period removed because it sits inside a longer sentence; {options} one option per line as `<letter>. <text>`, in stored order; {side} which wrist camera is being asked about; {goal} the scenario's overall goal. Text options are always rendered into the question (the three action dimensions and frame_order); image options are attached as media instead (frame_match, view_match). Without that rule a harness gets it wrong either way: showing `question` alone drops frame_order's options, and appending them itself prints the other three dimensions' options twice. The wording lives in the dataset rather than in evaluation code because changing it changes what the model was asked, which makes scores incomparable — that is something a dataset version has to record.",
+ "templates": {
+  "current_action": "Based on the egocentric video up to now, choose the ONE option that best matches what is happening RIGHT NOW?\nOptions:\n{options}",
+  "next_action": "Based on the current visual state, what should happen next?\nOptions:\n{options}",
+  "next_action_with_goal": "The overall task is {goal}. Based on the current visual state, what should happen next?\nOptions:\n{options}",
+  "frame_match": "Given this video clip of an action segment, which option image appeared in the clip?",
+  "view_match": "Given the image captured by the head camera, which option shows the {side} gripper camera's view at this moment?",
+  "frame_order": "The images above show three moments from the same episode, presented in random order. Which option lists them in the correct chronological order?\nOptions:\n{options}",
+  "action_time": "When did the action \"{subtask}\" happen?"
+ }
+}

tidy_stationery_gim/qa/scenarios.json

@@ -0,0 +1,31 @@
+{
+  "tidy_stationery_gim": {
+    "embodiment": "gim",
+    "goal": "capping a pen and putting it into its box",
+    "camera": {
+      "resolution": [
+        736,
+        416
+      ],
+      "native": null,
+      "fps": 50.0,
+      "views": [
+        "main",
+        "wrist_left",
+        "wrist_right"
+      ],
+      "timestamp_bar_rows_cropped": null,
+      "resolution_variants": 1
+    },
+    "next_action": {
+      "transitions": 216,
+      "unique_successor_share": 1.0,
+      "conditional_entropy": 0.0
+    },
+    "counts": {
+      "episodes": 108,
+      "segments": 324,
+      "subtasks": 3
+    }
+  }
+}

tidy_stationery_gim/qa/subtasks.json

@@ -0,0 +1,40 @@
+{
+  "tidy_stationery_gim": [
+    {
+      "id": "put_cap_on_pen",
+      "text": "Put the cap on the pen.",
+      "synthetic": false
+    },
+    {
+      "id": "place_pen_in_holder",
+      "text": "Place the pen in the holder.",
+      "synthetic": false
+    },
+    {
+      "id": "pick_cap_and_pen",
+      "text": "Pick the cap and pen.",
+      "synthetic": false
+    },
+    {
+      "id": "remove_cap_from_pen",
+      "text": "Remove the cap from the pen.",
+      "synthetic": true,
+      "absent_because": "The task puts the cap on the pen; removing the cap from the pen never occurs.",
+      "evidence": "A"
+    },
+    {
+      "id": "drop_pen_on_table",
+      "text": "Drop the pen on the table.",
+      "synthetic": true,
+      "absent_because": "The pen is picked up, capped, and placed in the holder; dropping it on the table never occurs.",
+      "evidence": "B"
+    },
+    {
+      "id": "push_holder_aside",
+      "text": "Push the holder aside.",
+      "synthetic": true,
+      "absent_because": "The pen is placed into the holder, but the holder itself is never pushed aside.",
+      "evidence": "B"
+    }
+  ]
+}