Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

ManipBench: Benchmarking Vision-Language Models for Low-Level Robot Manipulation (2505.09698v1)

Published 14 May 2025 in cs.RO and cs.AI

Abstract: Vision-LLMs (VLMs) have revolutionized artificial intelligence and robotics due to their commonsense reasoning capabilities. In robotic manipulation, VLMs are used primarily as high-level planners, but recent work has also studied their lower-level reasoning ability, which refers to making decisions about precise robot movements. However, the community currently lacks a clear and common benchmark that can evaluate how well VLMs can aid low-level reasoning in robotics. Consequently, we propose a novel benchmark, ManipBench, to evaluate the low-level robot manipulation reasoning capabilities of VLMs across various dimensions, including how well they understand object-object interactions and deformable object manipulation. We extensively test 33 representative VLMs across 10 model families on our benchmark, including variants to test different model sizes. Our evaluation shows that the performance of VLMs significantly varies across tasks, and there is a strong correlation between this performance and trends in our real-world manipulation tasks. It also shows that there remains a significant gap between these models and human-level understanding. See our website at: https://manipbench.github.io.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (8)
  1. Enyu Zhao (6 papers)
  2. Vedant Raval (3 papers)
  3. Hejia Zhang (24 papers)
  4. Jiageng Mao (20 papers)
  5. Zeyu Shangguan (11 papers)
  6. Stefanos Nikolaidis (65 papers)
  7. Yue Wang (676 papers)
  8. Daniel Seita (40 papers)

Summary

ManipBench: Benchmarking Vision-LLMs for Low-Level Robot Manipulation

The paper "ManipBench: Benchmarking Vision-LLMs for Low-Level Robot Manipulation" addresses a pivotal challenge in the intersection of vision-LLMs (VLMs) and robotics. VLMs, characterized by their commonsense reasoning capabilities, have predominantly been utilized for high-level planning applications in robotics. However, their utility in precise, low-level robot manipulation remains less explored, largely owing to the absence of benchmarks tailored for such tasks. The authors propose a benchmark, ManipBench, aimed at evaluating VLMs' ability to assist in low-level robotic manipulation, thus filling a critical gap in current research.

Key Contributions and Methodology:

  1. Benchmark Structure: ManipBench is an open-source benchmark that utilizes multiple-choice questions to assess VLMs' understanding of low-level robotic manipulation. It targets several dimensions, including object-object interactions and deformable object manipulation, without requiring trajectory rollouts. This method offers an efficient evaluation of reasoning capabilities by VLMs.
  2. Evaluation Dataset: The benchmark includes 12,617 questions derived from a diverse set of sources—existing public datasets, a custom in-house fabric manipulation setup, and simulation environments. This diversity ensures a comprehensive evaluation across tasks like pick-and-place, articulated object manipulation, and dynamic manipulation.
  3. Model Evaluation: The authors extensively test 33 representative VLMs across 10 model families, including both open-source and closed-source models like GPT-4 and Gemini. This evaluation highlights significant performance variability across tasks and identifies models such as Gemini-2.5-pro as top performers. The findings emphasize a considerable gap between VLM capabilities and human-level understanding in manipulation tasks.
  4. Real-World Correlation: The paper demonstrates a strong correlation between model performance on ManipBench and their real-world effectiveness, thereby validating the benchmark's utility as a proxy for evaluating VLMs in embodied robotic settings.
  5. Implications and Future Directions: The findings reveal that while the best-performing models can outperform random chance significantly, there remains substantial room for improvement in VLMs' reasoning capabilities in robotic manipulation. This highlights the need for continued innovation in model development.

Practical and Theoretical Implications:

The introduction of ManipBench has profound implications for both the development and evaluation of VLMs in robotics. Practically, it provides a framework for systematically assessing the effectiveness of various VLMs, guiding the selection of models for specific robotic manipulation tasks. Theoretically, the benchmark's insights into model performance variability offer valuable lessons for refining model architectures and training approaches.

Furthermore, the benchmark's design encourages a focus on low-level reasoning capabilities, vital for tasks requiring precise manipulation, such as deformable object handling—a challenging area that remains underexplored compared to rigid object manipulation. As the field progresses, leveraging VLMs to bridge the sim-to-real gap in robotics presents a promising avenue for research. ManipBench sets a foundational stepping stone in this journey.

In conclusion, the paper marks a significant contribution to robotics and artificial intelligence by addressing the evaluation gap for VLMs in low-level manipulation tasks. The benchmark not only provides a comprehensive tool for assessing current model capabilities but also encourages future advancements that push the boundary towards generalist robotic solutions.

Github Logo Streamline Icon: https://streamlinehq.com
Youtube Logo Streamline Icon: https://streamlinehq.com