Upload bench/protocol

bench/protocol/action.json

@@ -0,0 +1 @@
+{"format":"dream-exe.protocol","stage":"action","values":{"reference_frame":"world"}}

bench/protocol/evaluation.json

@@ -0,0 +1,69 @@
+{
+  "format": "dream-exe.protocol",
+  "stage": "evaluation",
+  "values": {
+    "vlm": {
+      "task_prompt_suffixes": {
+        "standard": "(Static surveillance camera. No camera movement. One continuous shot. The robot arm moves to manipulate the target object.)",
+        "enhanced": ""
+      },
+      "rubrics": {
+        "subject_stability": {
+          "prompt_files": {
+            "robot_subject": "vlm/prompts/subject_stability_robot.txt",
+            "manipulated_object": "vlm/prompts/subject_stability_object.txt"
+          },
+          "sampling": {
+            "frame_fractions": [0.0, 0.75],
+            "grid": [1, 2]
+          },
+          "score_range": [1, 15]
+        },
+        "physical_plausibility": {
+          "prompt_file": "vlm/prompts/physical_plausibility.txt",
+          "sampling": {
+            "uniform_frames": 6,
+            "grid": [3, 2]
+          },
+          "score_range": [1, 5]
+        },
+        "task_adherence": {
+          "prompt_file": "vlm/prompts/task_adherence.txt",
+          "sampling": {
+            "uniform_frames": 6,
+            "grid": [3, 2]
+          },
+          "score_range": [1, 5]
+        }
+      },
+      "judges": [
+        {
+          "id": "gemini-3-pro",
+          "provider": "openai-compatible",
+          "model": "gemini-3-pro",
+          "base_url_env": "DREAM_EXE_GEMINI_BASE_URL",
+          "api_key_env": "DREAM_EXE_GEMINI_API_KEY"
+        },
+        {
+          "id": "qwen3-vl-plus",
+          "provider": "openai-compatible",
+          "model": "qwen3-vl-plus",
+          "base_url_env": "DREAM_EXE_QWEN_BASE_URL",
+          "api_key_env": "DREAM_EXE_QWEN_API_KEY"
+        }
+      ],
+      "aggregation": {
+        "method": "mean",
+        "required_judges": ["gemini-3-pro", "qwen3-vl-plus"]
+      }
+    },
+    "trajectory": {
+      "groups": ["EEF vis", "EEF tcp", "OBJ"],
+      "metrics": ["HSD", "DYN", "NDTW"]
+    },
+    "execution": {
+      "executability_metrics": ["E-SR", "nDTW", "Pos95", "Rot95", "Smth"],
+      "task_metrics": ["SR-B", "SR-P", "Rel", "Place", "Art", "Core"]
+    }
+  }
+}

bench/protocol/execution.json

@@ -0,0 +1 @@
+{"format":"dream-exe.protocol","stage":"execution","values":{}}

bench/protocol/generation.json

@@ -0,0 +1 @@
+{"format":"dream-exe.protocol","stage":"generation","values":{"first_frame_required":true,"kind":"image_to_video","prompt_variants":["standard","enhanced"]}}

bench/protocol/video2traj.json

@@ -0,0 +1 @@
+{"format":"dream-exe.protocol","stage":"video2traj","values":{"pose":{"backend":"foundationpose","config_path":""},"region":{"targets":{"eef":{"simulation":{"instance_name":"gripper0_right_right_gripper"}}}}}}

bench/protocol/vlm/prompts/physical_plausibility.txt

@@ -0,0 +1,20 @@
+The provided image presents sequential frames, arranged in a grid, from a {view} perspective AI-generated task video about {description}
+Does this video comply with common-sense expectations for human-level interactions?
+
+# A. Anomaly Checks:
+1) Physical grounding violation: any part of the robot appears floating, or intersecting/penetrating other geometry.
+2) Spontaneous object appearance: any object or robotic agent that suddenly appears between frames without a plausible cause. Pay special attention to the top two frames, where such anomalies often occur early in the sequence.
+3) Non-contact attachment / false grasp: If the video involves a grasping interaction, carefully check whether the object remains attached to or moves with the robotic gripper/hand without clear physical contact, closure.
+If either anomaly is present, treat it as a major violation and assign a low score (1–2).
+
+# B. Human-Level Common-Sense Consistency:
+Evaluate whether this video complies with common-sense expectations for human-level interactions?
+Rate the video on a scale from 1 to 5, where 5 means fully consistent with human common sense and 1 means major violations.
+
+### Special Notes:
+1) Be very cautious when assigning scores of 4 or 5: these scores should only be used when clearly correct, and free of errors. Do not give 4 or 5 lightly.
+2) Use step-by-step reasoning internally to make your selection.
+3) Your output must be a valid JSON object with two fields:
+- "reason": a brief justification for the given score
+- "score": an integer between 1 and 5
+Return exactly: {json_example}

bench/protocol/vlm/prompts/subject_stability_object.txt

@@ -0,0 +1,16 @@
+The provided image shows two sequential frames from an AI-generated video about robot doing a task.
+The left frame is the correct reference image, while the right frame is the AI-generated video frame.
+Focus on how '{manipulated_object}' appears in both frames, and evaluate the consistency of '{manipulated_object}' between the reference and the generated frame.
+
+Note:
+1) Changes in orientation or position are acceptable and should not affect the consistency rating.
+2) Important: Do NOT assign option A or B lightly.
+
+Question:
+A: '{manipulated_object}' in the right frame is clear and consistent with the left image.
+B: '{manipulated_object}' in the right frame is mostly consistent with the left image, with minor visual issues.
+C: '{manipulated_object}' in the right frame shows noticeable inconsistencies compared with the left image.
+D: '{manipulated_object}' in the right frame undergoes a major transformation, appears as an AI-generated artifact or is duplicated compared with the left image.
+E: '{manipulated_object}' is missing in the right frame while it exists in the left image.
+The options A to E represent increasing levels of inconsistency. Select the most suitable option.
+Return exactly: {json_example}

bench/protocol/vlm/prompts/subject_stability_robot.txt

@@ -0,0 +1,17 @@
+The provided image shows two sequential frames from an AI-generated video about robot doing a task.
+The left frame is the correct reference image, while the right frame is the AI-generated video frame.
+Focus on how '{robot_subject}' appears in both frames, and evaluate the consistency of '{robot_subject}' between the reference and the generated frame.
+
+Note:
+1) Pay special attention to distinguishing between robotic gripper and robotic hand (if visible). Robotic gripper usually has a small number of rigid gripping jaws or prongs, while a robotic hand has multiple articulated fingers and more complex structures.
+2) Changes in orientation or position are acceptable and should not affect the consistency rating.
+3) Important: Do NOT assign option A or B lightly.
+
+Question:
+A: '{robot_subject}' in the right frame is clear and consistent with the left image.
+B: '{robot_subject}' in the right frame is mostly consistent with the left image, with minor visual issues.
+C: '{robot_subject}' in the right frame shows noticeable inconsistencies compared with the left image, such as changes in shape or structure.
+D: '{robot_subject}' in the right frame is highly inconsistent with the left image or transforms into another type of '{robot_subject}'.
+E: {terminal_option}
+The options A to E represent increasing levels of inconsistency. Select the most suitable option.
+Return exactly: {json_example}

bench/protocol/vlm/prompts/task_adherence.txt

@@ -0,0 +1,11 @@
+The provided image presents sequential frames, arranged in a grid, from a {view} perspective AI-generated task video.
+In this AI-generated video, does the robot successfully perform the task: "{description}"?
+Please rate the video on a scale from 1 to 5, where 5 indicates a perfect match and 1 indicates no relevance.
+
+### Special Notes:
+1) Be very cautious when assigning scores of 4 or 5: these scores should only be used when clearly correct, and free of errors. Do not give 4 or 5 lightly.
+2) Use step-by-step reasoning internally to make your selection.
+3) Your output must be a valid JSON object with two fields:
+- "reason": a brief justification for the given score
+- "score": an integer between 1 and 5
+Return exactly: {json_example}