RoboVista: Evaluating Vision-Language Models for Diverse Robot Applications

RSS 2026 · Project page · Leaderboard · Code

RoboVista is a benchmark of 474 multiple-choice questions over real robot scenes for evaluating vision-language models on the perception and reasoning skills robots actually need. Each question pairs one or more images from a real robot deployment (wrist cameras, overhead views, driving scenes, surgical setups) with 2–5 answer choices, a ground-truth answer, and expert reasoning.

Questions span six application domains — agriculture, autonomous driving, domestic, industrial manufacturing, surgical robotics, and scenes from open robot-learning datasets (DROID, Bridge, AgiBot, Fractal, Dex-Net, …) — and are labeled by ability type: scene understanding (geometry & spatial reasoning, sequential events, physical interaction) vs. planning & decision-making (goal/action reasoning, motion feasibility, failure recovery).

Dataset structure

Field	Type	Description
images	list[Image]	1–7 images per question
publication_source	string	Source dataset / paper for the scene
question	string	Question text
choices	list[string]	Up to 5 answer options (A–E; unused slots empty)
correct_answer	string	Ground-truth letter (A–E)
reasoning	string	Annotator's reasoning for the answer
id	string	Question id
domain	string	Application domain (see above)
task	string	Original task description from the source deployment
ability_type	string	scene_understanding, high_level_decision_making, low_level_motion_awareness, or recovery_replanning_robustness
ability_subcategory	string	Fine-grained ability label
creator_name	string	Annotator

Single train split, 474 rows, images embedded (~1.1 GB).

Usage

from datasets import load_dataset

ds = load_dataset("sy-xie/robovista", split="train")
ex = ds[0]

letters = ["A", "B", "C", "D", "E"]
choices = "\n".join(
    f"{letter}: {c}" for letter, c in zip(letters, ex["choices"]) if c
)
prompt = (
    f"{ex['question']}\n\n{choices}\n\n"
    "Answer with the letter only (A, B, C, D, or E)."
)
# send ex["images"] + prompt to your VLM, compare to ex["correct_answer"]

The evaluation harness (server launch scripts, standard/CoT prompts, answer parsing) is in the code repository.

Leaderboard

Accuracy with the standard prompt and a chain-of-thought prompt (top models shown; full results and per-domain breakdowns on the leaderboard page):

Model	Weights	Standard	CoT
Gemini 3 Flash (preview)	API	68.9	65.4
Qwen3.5-397B-A17B	open	55.2	52.8
GPT-5-0	API	54.6	54.6
Qwen3.6-27B	open	53.5	54.4
Gemma 4 31B	open	51.5	55.4
Qwen3.6-35B-A3B	open	50.9	50.0
Qwen3-VL-32B-Thinking	open	49.6	51.3
GPT-4o	API	48.5	—
GLM-4.6V	open	48.0	45.0
RoboBrain2.5-8B-NV	open	45.7	40.4
Cosmos-Reason2-32B	open	46.1	42.0
Cosmos3-Nano (16B)	open	44.4	39.8
Gemma 4 12B	open	41.7	45.4
Molmo2-8B	open	41.7	33.0
Cosmos-Reason1-7B	open	41.3	39.8
Cosmos-Reason2-2B	open	38.3	33.5

Scoring notes: 460 of 474 questions are scored (a 14-question blacklist is excluded); answers are extracted with the harness's improved parser; Qwen3/3.5/3.6 rows run in non-thinking mode.

Citation

@inproceedings{robovista2026,
  title     = {RoboVista: Evaluating Vision-Language Models for Diverse Robot Applications},
  author    = {Xie, Shuangyu and Chen, Kaiyuan and Chen, Ziyang and Adebola, Simeon and
               Huang, Yixuan and Ma, Zehan and Qiu, Tianshuang and Yuan, Wentao and
               Shah, Dhruv and Sanketi, Pannag R. and Goldberg, Ken},
  booktitle = {Robotics: Science and Systems (RSS)},
  year      = {2026}
}