diff --git a/lmms_eval/tasks/robo_spatial/_default_template.yaml b/lmms_eval/tasks/robo_spatial/_default_template.yaml
new file mode 100644
index 000000000..8d64c0b4f
--- /dev/null
+++ b/lmms_eval/tasks/robo_spatial/_default_template.yaml
@@ -0,0 +1,13 @@
+dataset_path: chanhee-luke/RoboSpatial-Home
+output_type: generate_until
+process_docs: !function utils.robo_spatial_process_docs
+doc_to_visual: !function utils.robo_spatial_doc_to_visual
+doc_to_text: !function utils.robo_spatial_doc_to_text
+doc_to_target: !function utils.robo_spatial_doc_to_answer
+process_results: !function utils.robo_spatial_process_results_generation
+metric_list:
+  - metric: robo_spatial_score
+    aggregation: !function utils.robo_spatial_aggregate_score
+    higher_is_better: true
+metadata:
+  version: 0.1
diff --git a/lmms_eval/tasks/robo_spatial/pre_process.py b/lmms_eval/tasks/robo_spatial/pre_process.py
new file mode 100644
index 000000000..2a603dcbc
--- /dev/null
+++ b/lmms_eval/tasks/robo_spatial/pre_process.py
@@ -0,0 +1,188 @@
+### reference: https://github.com/flageval-baai/FlagEvalMM/blob/main/flagevalmm/evaluator/pre_process.py
+
+import re
+
+
+def strip_answer(answer):
+    answer = re.sub("The", "", answer)
+    answer = re.sub("If", "", answer)
+    answer = re.sub("[INST]", "", answer)
+    answer = re.sub("[/INST]", "", answer)
+    answer = re.sub("<Img>", "", answer)
+    answer = re.sub("</Img>", "", answer)
+    answer = answer.strip()
+    return answer
+
+
+def remove_special_characters(text):
+    pattern = (
+        r"[-`\\【】\*\$、,，。.；;:：？\?！!\s\n\u4e00-\u9fff0-9①②③④⑤⑥⑦\[\]\<>a-z=\'\"\(\)\{\}]+"
+    )
+    cleaned_text = re.sub(pattern, "", text)
+
+    return cleaned_text
+
+
+def process_multiple_choice(answer):
+    answer = strip_answer(answer)
+    pattern = r"^([A-Z])\."
+    matches = re.match(pattern, answer)
+    if matches:
+        return matches.group(1)
+    key_words = [
+        "boxed",
+        "Answer:",
+        "Answer is",
+        "answer is",
+        "option is",
+        "Correct option",
+        "correct option",
+        "Answer",
+        "answer",
+        "故选",
+        "选择",
+        "正确选项为",
+        "答案选",
+        "答案为",
+        "答案是",
+        "因此",
+        "答案",
+    ]
+
+    for key_word in key_words:
+        if key_word in answer:
+            answer = answer.split(key_word)[-1]
+            break
+    answer = remove_special_characters(answer)
+    # keep the last line
+    answer = answer.split("\n")[-1]
+    pattern = r"[A-Z]"
+    matches = re.findall(pattern, answer)
+    return "".join(matches)
+
+
+def remove_unit(value):
+    units = [
+        "cm",
+        "m",
+        "km",
+        "mm",
+        "s",
+        "h",
+        "kg",
+        "g",
+        "l",
+        "ml",
+        "mol",
+        "厘米",
+        "米",
+        "千米",
+        "°",
+        "毫米",
+        "月",
+        "秒",
+        "小时",
+        "克",
+        "千克",
+        "升",
+        "毫升",
+        "摩尔",
+    ]
+    unit_pattern = r"^(\d+)(?:" + "|".join(units) + ")$"
+    match = re.match(unit_pattern, value)
+    if match:
+        return match.group(1)
+    else:
+        return value
+
+
+def convert_circled_numbers(text):
+    circled_numbers = {
+        "①": "1",
+        "②": "2",
+        "③": "3",
+        "④": "4",
+        "⑤": "5",
+        "⑥": "6",
+        "⑦": "7",
+        "⑧": "8",
+        "⑨": "9",
+        "⑩": "10",
+    }
+    for circled, number in circled_numbers.items():
+        text = text.replace(circled, number)
+    return text
+
+
+def normalize_string(raw_answer):
+    if "$" not in raw_answer:
+        wrong_answer_words = ["\\times", "不对", "不正确", "×"]
+        for word in wrong_answer_words:
+            raw_answer = raw_answer.replace(word, "错误")
+    raw_answer = re.sub(r"\\text\s*\{(.*?)\}", r"\1", raw_answer)
+    replace_dict = {
+        "√": "正确",
+        "：": ":",
+        "$": "",
+        "（": "(",
+        "）": ")",
+        "，": ",",
+        "。": ".",
+        "变小": "减小",
+        "变大": "增大",
+        "路程": "距离",
+        "\\pi": "π",
+        "＞": ">",
+        "＜": "<",
+        "；": ";",
+    }
+    # write to convert characters like ①②③④ to 1234
+
+    for k, v in replace_dict.items():
+        raw_answer = raw_answer.replace(k, v)
+
+    # Convert circled numbers to regular numbers
+    raw_answer = convert_circled_numbers(raw_answer)
+
+    strict_replace_dict = {
+        "错": "错误",
+        "对": "正确",
+        "(F)": "F",
+        "(T)": "T",
+        "(正确)": "正确",
+        "(错误)": "错误",
+        "“T”": "T",
+        "“F”": "F",
+    }
+    if raw_answer in strict_replace_dict:
+        raw_answer = strict_replace_dict[raw_answer]
+
+    key_words = [
+        "Answer:",
+        "Answer is",
+        "answer is",
+        "Answer",
+        "answer",
+        "答案为",
+        "答案是",
+        "解是",
+        "解为",
+        "答案",
+        "结果",
+        "为",
+        "因此",
+        " = ",
+    ]
+    # get text after key_word
+    for key_word in key_words:
+        if key_word in raw_answer:
+            raw_answer = raw_answer.split(key_word)[-1]
+            break
+    raw_answer = raw_answer.strip()
+    # remove leading :
+    if raw_answer.startswith(":"):
+        raw_answer = raw_answer[1:]
+    if len(raw_answer) > 0 and raw_answer[-1] in [".", ",", ":", ";"]:
+        raw_answer = raw_answer[:-1]
+    raw_answer = remove_unit(raw_answer)
+    return raw_answer.strip()
diff --git a/lmms_eval/tasks/robo_spatial/robo_spatial.yaml b/lmms_eval/tasks/robo_spatial/robo_spatial.yaml
new file mode 100644
index 000000000..ac32a1f00
--- /dev/null
+++ b/lmms_eval/tasks/robo_spatial/robo_spatial.yaml
@@ -0,0 +1,3 @@
+group: robo_spatial
+task:
+  - robo_spatial_all
diff --git a/lmms_eval/tasks/robo_spatial/robo_spatial_all.yaml b/lmms_eval/tasks/robo_spatial/robo_spatial_all.yaml
new file mode 100644
index 000000000..bc6eb7382
--- /dev/null
+++ b/lmms_eval/tasks/robo_spatial/robo_spatial_all.yaml
@@ -0,0 +1,24 @@
+task: robo_spatial_all
+test_split: combined
+
+dataset_path: chanhee-luke/RoboSpatial-Home
+dataset_kwargs:
+  data_files:
+    combined:
+      - data/compatibility-*.parquet
+      - data/configuration-*.parquet
+      - data/context-*.parquet
+  verification_mode: no_checks
+
+output_type: generate_until
+process_docs: !function utils.robo_spatial_process_docs
+doc_to_visual: !function utils.robo_spatial_doc_to_visual
+doc_to_text: !function utils.robo_spatial_doc_to_text
+doc_to_target: !function utils.robo_spatial_doc_to_answer
+process_results: !function utils.robo_spatial_process_results_generation
+metric_list:
+  - metric: robo_spatial_score
+    aggregation: !function utils.robo_spatial_aggregate_score
+    higher_is_better: true
+metadata:
+  version: 0.1
diff --git a/lmms_eval/tasks/robo_spatial/robo_spatial_compatibility.yaml b/lmms_eval/tasks/robo_spatial/robo_spatial_compatibility.yaml
new file mode 100644
index 000000000..fee479428
--- /dev/null
+++ b/lmms_eval/tasks/robo_spatial/robo_spatial_compatibility.yaml
@@ -0,0 +1,3 @@
+task: robo_spatial_compatibility
+include: _default_template.yaml
+test_split: compatibility
diff --git a/lmms_eval/tasks/robo_spatial/robo_spatial_configuration.yaml b/lmms_eval/tasks/robo_spatial/robo_spatial_configuration.yaml
new file mode 100644
index 000000000..c4a643a43
--- /dev/null
+++ b/lmms_eval/tasks/robo_spatial/robo_spatial_configuration.yaml
@@ -0,0 +1,3 @@
+task: robo_spatial_configuration
+include: _default_template.yaml
+test_split: configuration
diff --git a/lmms_eval/tasks/robo_spatial/robo_spatial_context.yaml b/lmms_eval/tasks/robo_spatial/robo_spatial_context.yaml
new file mode 100644
index 000000000..0a5da1914
--- /dev/null
+++ b/lmms_eval/tasks/robo_spatial/robo_spatial_context.yaml
@@ -0,0 +1,3 @@
+task: robo_spatial_context
+include: _default_template.yaml
+test_split: context
diff --git a/lmms_eval/tasks/robo_spatial/utils.py b/lmms_eval/tasks/robo_spatial/utils.py
new file mode 100644
index 000000000..9e9341b61
--- /dev/null
+++ b/lmms_eval/tasks/robo_spatial/utils.py
@@ -0,0 +1,262 @@
+import hashlib
+import os
+import re
+import tempfile
+
+import numpy as np
+import pandas as pd
+from PIL import Image
+
+_MASK_CACHE_DIR = os.path.join(tempfile.gettempdir(), "robo_spatial_masks")
+os.makedirs(_MASK_CACHE_DIR, exist_ok=True)
+
+
+def remove_think(content: str) -> str:
+    content = re.sub(r"^.*?</think>", "", content, flags=re.DOTALL)
+    content = re.sub(r"^.*?</thought>", "", content, flags=re.DOTALL)
+    match = re.search(r"<answer>(.*?)</answer>", content, flags=re.DOTALL | re.IGNORECASE)
+    if match:
+        return match.group(1).strip()
+    return content.strip()
+
+
+def json2pts(text: str, width=640, height=480) -> np.ndarray:
+    import json
+
+    match = re.search(r"```(?:\w+)?\n(.*?)```", text, re.DOTALL)
+    if not match:
+        return np.empty((0, 2), dtype=int)
+
+    try:
+        data = json.loads(match.group(1).strip())
+    except json.JSONDecodeError:
+        return np.empty((0, 2), dtype=int)
+
+    points = []
+    for item in data:
+        if "point_2d" in item and isinstance(item["point_2d"], list) and len(item["point_2d"]) == 2:
+            x_norm, y_norm = item["point_2d"]
+            x = int(x_norm / 1000 * width)
+            y = int(y_norm / 1000 * height)
+            points.append((x, y))
+    return np.array(points)
+
+
+def evaluate_binary_answer(ground_truth, generated_answer):
+    gen_answer = generated_answer.strip().lower().split("\n")[-1]
+    is_gt_yes = ground_truth.strip().lower() == "yes"
+    if is_gt_yes:
+        correct = "yes" in gen_answer and "no" not in gen_answer
+    else:
+        correct = "no" in gen_answer and "yes" not in gen_answer
+    return correct
+
+
+def _save_mask(doc):
+    from io import BytesIO
+
+    mask = doc.get("mask")
+    if mask is None:
+        return {"mask_path": ""}
+    # HuggingFace datasets stores image columns as {'bytes': b'...'} dicts
+    # when the column is not declared with Image() feature — decode it here.
+    if isinstance(mask, dict) and mask.get("bytes"):
+        try:
+            mask = Image.open(BytesIO(mask["bytes"]))
+        except Exception:
+            return {"mask_path": ""}
+    if isinstance(mask, Image.Image):
+        doc_hash = hashlib.md5(doc.get("question", "").encode()).hexdigest()
+        mask_path = os.path.join(_MASK_CACHE_DIR, f"{doc_hash}.png")
+        if not os.path.exists(mask_path):
+            mask.save(mask_path)
+        return {"mask_path": mask_path}
+    return {"mask_path": ""}
+
+
+def robo_spatial_process_docs(dataset):
+    # load_from_cache_file=False re-saves mask PNGs to /tmp on every run,
+    # preventing stale cached mask_path values pointing to deleted temp files.
+    return dataset.map(_save_mask, load_from_cache_file=False)
+
+
+def evaluate_pointing_with_mask(doc, generated_answer):
+    mask_path = doc.get("mask_path")
+    if not mask_path or not os.path.exists(mask_path):
+        return -1.0
+
+    try:
+        mask = np.array(Image.open(mask_path))
+        if mask.ndim == 3:
+            mask = mask[:, :, 0]
+        mask = (mask > 0).astype(np.uint8)
+    except Exception:
+        return -1.0
+
+    h, w = mask.shape
+    pred_points = json2pts(generated_answer, width=w, height=h)
+    if len(pred_points) == 0:
+        return 0.0
+
+    for pt in pred_points:
+        x, y = int(pt[0]), int(pt[1])
+        if 0 <= x < w and 0 <= y < h and mask[y, x]:
+            return 1.0
+    return 0.0
+
+
+def evaluate_pointing_with_mask_first_only(doc, generated_answer):
+    # Strict: only the first predicted point counts. Removes the "any point in
+    # mask" loophole — the model must commit to a single best point.
+    mask_path = doc.get("mask_path")
+    if not mask_path or not os.path.exists(mask_path):
+        return -1.0
+
+    try:
+        mask = np.array(Image.open(mask_path))
+        if mask.ndim == 3:
+            mask = mask[:, :, 0]
+        mask = (mask > 0).astype(np.uint8)
+    except Exception:
+        return -1.0
+
+    h, w = mask.shape
+    pred_points = json2pts(generated_answer, width=w, height=h)
+    if len(pred_points) == 0:
+        return 0.0
+    x, y = int(pred_points[0][0]), int(pred_points[0][1])
+    if 0 <= x < w and 0 <= y < h and mask[y, x]:
+        return 1.0
+    return 0.0
+
+
+def robo_spatial_doc_to_visual(doc, lmms_eval_specific_kwargs=None):
+    return [doc["img"]]
+
+
+_HIRES_XL_TARGET_PIXELS = 1536 * 28 * 28  # 1,204,224 px ≈ 1536 vision tokens
+
+
+def _upscale_if_small(img, target_pixels):
+    from PIL import Image as _PILImage
+
+    w, h = img.size
+    cur = w * h
+    if cur >= target_pixels:
+        return img
+    scale = (target_pixels / cur) ** 0.5
+    return img.resize((int(round(w * scale)), int(round(h * scale))), _PILImage.BICUBIC)
+
+
+def robo_spatial_doc_to_visual_hires_xl_pointing_only(doc, lmms_eval_specific_kwargs=None):
+    # Upscale only pointing (context) images — yes/no categories are unaffected
+    # by visual detail and over-upscaling hurts them.
+    img = doc["img"]
+    if img.mode != "RGB":
+        img = img.convert("RGB")
+    if doc.get("category") == "context":
+        return [_upscale_if_small(img, _HIRES_XL_TARGET_PIXELS)]
+    return [img]
+
+
+def get_post_prompt(doc):
+    # reference: https://github.com/flageval-baai/FlagEvalMM/blob/main/tasks/robo_spatial_home/robo_spatial_home_all.py
+    post_prompt_point = """Your task is to identify specific points in the image based on the question. Respond with a brief explanation if needed, followed by a list of 2D point coordinates.
+
+    The points should be in JSON format and normalized to the range [0, 1000].
+
+    Do not include additional text after this line.
+    """
+
+    post_prompt_yes_no = """Your task is to answer the question above. Respond with a brief explanation if needed, followed by a yes or no answer in the last line of your response.
+
+    Format your final answer strictly as follows on the last line of your response:
+    Answer: yes or no
+
+    Do not include additional text after this line.
+    """
+    if doc["category"] == "context":
+        return post_prompt_point
+    else:
+        return post_prompt_yes_no
+
+
+def get_post_prompt_one_point_v2(doc):
+    # Single-point, terse variant emphasizing "NOT on the object".
+    post_prompt_point = """Pinpoint the SINGLE best 2D point in the VACANT SPACE described by the question. The point MUST NOT lie on the reference object. It MUST lie in empty space on the specified side.
+
+    Output exactly one coordinate pair in JSON, normalized to [0, 1000]:
+    ```json
+    [{"point_2d": [x, y], "label": "vacant space"}]
+    ```
+
+    Do not output more than one point. Do not include additional text after the JSON block.
+    """
+    post_prompt_yes_no = """Your task is to answer the question above. Respond with a brief explanation if needed, followed by a yes or no answer in the last line of your response.
+
+    Format your final answer strictly as follows on the last line of your response:
+    Answer: yes or no
+
+    Do not include additional text after this line.
+    """
+    if doc["category"] == "context":
+        return post_prompt_point
+    else:
+        return post_prompt_yes_no
+
+
+_QUESTION_SUFFIX_STRIP = [
+    "Your answer should be formatted as a list of tuples, i.e. [(x1, y1), ...], where each tuple contains the x and y coordinates of a point satisfying the conditions above. The coordinates should be between 0 and 1, indicating the normalized pixel locations of the points.",
+    "Answer yes or no.",
+]
+
+
+def _clean_question(doc):
+    question = doc.get("question", "")
+    for s in _QUESTION_SUFFIX_STRIP:
+        question = question.replace(s, "").strip()
+    return question
+
+
+def robo_spatial_doc_to_text(doc, lmms_eval_specific_kwargs=None):
+    return f"{_clean_question(doc)}\n{get_post_prompt(doc)}"
+
+
+def robo_spatial_doc_to_text_one_point_v2(doc, lmms_eval_specific_kwargs=None):
+    return f"{_clean_question(doc)}\n{get_post_prompt_one_point_v2(doc)}"
+
+
+def robo_spatial_doc_to_answer(doc):
+    return doc.get("answer", "")
+
+
+def robo_spatial_process_results_generation(doc, result):
+    pred = remove_think(result[0])
+    ground_truth = robo_spatial_doc_to_answer(doc)
+    gt_lower = ground_truth.strip().lower()
+
+    if gt_lower in ["yes", "no"]:
+        score = int(evaluate_binary_answer(ground_truth, pred))
+    else:
+        score = evaluate_pointing_with_mask(doc, pred)
+
+    return {"robo_spatial_score": {"score": score, "category": doc["category"]}, "score": score}
+
+
+def robo_spatial_process_results_first_point(doc, result):
+    # Strict pointing scorer: only the first predicted point is checked.
+    pred = remove_think(result[0])
+    ground_truth = robo_spatial_doc_to_answer(doc)
+    gt_lower = ground_truth.strip().lower()
+    if gt_lower in ["yes", "no"]:
+        score = int(evaluate_binary_answer(ground_truth, pred))
+    else:
+        score = evaluate_pointing_with_mask_first_only(doc, pred)
+    return {"robo_spatial_score": {"score": score, "category": doc["category"]}, "score": score}
+
+
+def robo_spatial_aggregate_score(results, args):
+    df = pd.DataFrame(results)
+    overall_score = df["score"].mean()
+    category_score_dict = df.groupby("category")["score"].mean().to_dict()
+    return {"overall": overall_score, **category_score_dict}