"""ERQA (Embodied Reasoning QA, Gemini Robotics team, 2025): 400 multiple-choice questions over robot images (spatial, trajectory, action, state, task, multi-view reasoning, pointing). Data: HF mirror `FlagEval/ERQA` (converted from the official TFRecord). Images are interleaved into the question text at `visual_indices` (character offsets), as in the official evaluator. """ from __future__ import annotations import ast import io import re from pathlib import Path import numpy as np LETTERS = "ABCDEFGH" def load(path="data/erqa/test.parquet"): import pandas as pd from PIL import Image df = pd.read_parquet(path) items = [] for _, r in df.iterrows(): imgs = [Image.open(io.BytesIO(im["bytes"])).convert("RGB") for im in r["images"]] vis = list(r["visual_indices"]) if r["visual_indices"] is not None else [0] * len(imgs) if isinstance(vis, str): vis = ast.literal_eval(vis) opts = options(r["question"]) items.append({"id": r["question_id"], "question": r["question"], "type": r["question_type"], "answer": r["answer"].strip().upper()[:1], "images": imgs, "visual_indices": [int(v) for v in vis], "options": opts}) return items def options(question: str) -> str: """Letters offered in the question ('Choices: A. ... B. ...'); defaults to A-D.""" found = re.findall(r"(?:^|\s)([A-H])[\.\)]\s", question.split("Choices:")[-1] if "Choices:" in question else question) letters = "".join(sorted(set(found))) if len(letters) < 2: return "ABCD" return LETTERS[: LETTERS.index(letters[-1]) + 1] # fill gaps, e.g. "A. B. No. C. Yes." -> ABC(D) def interleave(item, suffix: str): """Chat content list with images inserted at their character offsets (official ERQA layout).""" q = item["question"] content, last = [], 0 for img, idx in sorted(zip(item["images"], item["visual_indices"]), key=lambda t: t[1]): idx = max(0, min(int(idx), len(q))) if q[last:idx].strip(): content.append({"type": "text", "text": q[last:idx]}) content.append({"type": "image", "image": img}) last = idx content.append({"type": "text", "text": q[last:] + suffix}) return content PROMPT_SUFFIX = "\nAnswer with the option's letter from the given choices directly." def accuracy_by_type(rows): out = {} for t in sorted({r["type"] for r in rows}): rs = [r for r in rows if r["type"] == t] out[t] = float(np.mean([r["pred"] == r["answer"] for r in rs])) out["overall"] = float(np.mean([r["pred"] == r["answer"] for r in rows])) return out