yay - works well!

Author: NishanthJKumar

.gitignore

@@ -42,3 +42,6 @@ top-down-state-view.png
 # Jetbrains IDEs
 .idea/
 hand_image_failed_detection.png
+
+# Spot wiping skill
+wipe_online_images_iphone/*

predicators/pretrained_model_interface.py

@@ -303,7 +303,7 @@ def _sample_completions(
                              candidate_count=num_completions)
         if self._supports_thinking():
             config_kwargs["thinking_config"] = types.ThinkingConfig(
-                thinking_budget=0)
+                thinking_budget=128)
         config = types.GenerateContentConfig(**config_kwargs)
         with _GeminiTimeout(seconds=30):
             response = self._client.models.generate_content(
@@ -339,7 +339,7 @@ def _sample_completions(
                              candidate_count=num_completions)
         if self._supports_thinking():
             config_kwargs["thinking_config"] = types.ThinkingConfig(
-                thinking_budget=0)
+                thinking_budget=128)
         config = types.GenerateContentConfig(**config_kwargs)
         with _GeminiTimeout(seconds=30):
             response = self._client.models.generate_content(

predicators/spot_utils/iphone_extrinsics.json

@@ -0,0 +1,56 @@
+{
+  "frame_convention": "T_hand_iphone (hand camera <- iphone camera)",
+  "translation_m": [
+    -0.013389742324134,
+    -0.09071333609668566,
+    -0.04515229560908661
+  ],
+  "rotation_matrix": [
+    [
+      0.9998152515149991,
+      0.0051358662617075794,
+      -0.018522573140379982
+    ],
+    [
+      -0.0023931047569163704,
+      0.9894048892806945,
+      0.1451627986678386
+    ],
+    [
+      0.01907186114728439,
+      -0.14509165360281362,
+      0.9892343888923292
+    ]
+  ],
+  "T_hand_iphone": [
+    [
+      0.9998152515149991,
+      0.0051358662617075794,
+      -0.018522573140379982,
+      -0.013389742324134
+    ],
+    [
+      -0.0023931047569163704,
+      0.9894048892806945,
+      0.1451627986678386,
+      -0.09071333609668566
+    ],
+    [
+      0.01907186114728439,
+      -0.14509165360281362,
+      0.9892343888923292,
+      -0.04515229560908661
+    ],
+    [
+      0.0,
+      0.0,
+      0.0,
+      1.0
+    ]
+  ],
+  "std_inliers_m": [
+    0.0013296844145321853,
+    0.0015158110809706138,
+    0.0005005343158977485
+  ]
+}

predicators/spot_utils/skills/spot_wipe_online_iphone.py

@@ -32,8 +32,8 @@
     move_hand_to_relative_pose,
     move_hand_to_relative_pose_with_velocity,
     open_gripper,
-    stow_arm,
 )
+from predicators.spot_utils.skills.spot_stow_arm import stow_arm
 from predicators.spot_utils.perception.spot_cameras import \
     _image_response_to_image
 from predicators.pretrained_model_interface import GoogleGeminiVLM
@@ -645,12 +645,9 @@ def get_bbox_from_gemini(
         List of [ymin, xmin, ymax, xmax] in pixel coordinates
 
     """
-    # Ensure API key is set for Gemini
-    # vlm = GoogleGeminiVLM("gemini-2.5-flash-preview-05-20")
+    from predicators.settings import CFG
     print('inside the function to get the bbox from gemini')
-    # vlm = GoogleGeminiVLM("gemini-2.5-flash")
-    # vlm = GoogleGeminiVLM("gemini-2.0-flash")
-    vlm = GoogleGeminiVLM("gemini-2.5-pro")
+    vlm = GoogleGeminiVLM(CFG.vlm_model_name)
     def _parse_bbox_list(raw: str) -> list[float]:
         """Parse a bbox dict {"bbox": [ymin, xmin, ymax, xmax]} from model output.
         Supports optional ```json fenced blocks. Returns raw numeric values
@@ -1055,7 +1052,18 @@ def main() -> None:
         default=0.0,
         help="Fraction to expand bbox in image space (e.g., 0.2 for +20%).",
     )
+    parser.add_argument(
+        "--vlm_model_name",
+        type=str,
+        default="gemini-2.5-flash",
+        help="Gemini model name to use for VLM queries.",
+    )
     args = parser.parse_args()
+
+    # Update CFG so get_bbox_from_gemini picks up the model name.
+    from predicators.settings import CFG
+    CFG.vlm_model_name = args.vlm_model_name
+
     robot, lease_client, lease_keepalive = init_robot(args.hostname, "")
 
     wipe_online(