Benchmarking Vision, Language, & Action Models on Robotic Learning Tasks

Pranav Guruprasad,Harshvardhan Sikka,Jaewoo Song,Yangyue Wang,Paul Pu Liang

2024-11-05

Abstract:Vision-language-action (VLA) models represent a promising direction for developing general-purpose robotic systems, demonstrating the ability to combine visual understanding, language comprehension, and action generation. However, systematic evaluation of these models across diverse robotic tasks remains limited. In this work, we present a comprehensive evaluation framework and benchmark suite for assessing VLA models. We profile three state-of-the-art VLM and VLAs - GPT-4o, OpenVLA, and JAT - across 20 diverse datasets from the Open-X-Embodiment collection, evaluating their performance on various manipulation tasks. Our analysis reveals several key insights: 1. current VLA models show significant variation in performance across different tasks and robot platforms, with GPT-4o demonstrating the most consistent performance through sophisticated prompt engineering, 2. all models struggle with complex manipulation tasks requiring multi-step planning, and 3. model performance is notably sensitive to action space characteristics and environmental factors. We release our evaluation framework and findings to facilitate systematic assessment of future VLA models and identify critical areas for improvement in the development of general purpose robotic systems.

Robotics,Computer Vision and Pattern Recognition,Machine Learning

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is that there are significant differences in the performance of current Vision - Language - Action (VLA) models on different robotic tasks and platforms, especially the lack of ability in complex manipulation tasks and multi - step planning. Specifically, the paper focuses on the following key issues:

1. **Performance differences**: Current VLA models show significant performance differences on different tasks and robotic platforms, and a systematic evaluation framework is required to fully understand the capabilities and limitations of these models.

2. **Complex - task - handling ability**: All models perform poorly when handling complex manipulation tasks that require multi - step planning, which limits their universality and reliability in practical applications.

3. **Sensitivity to environmental factors**: Model performance is very sensitive to the characteristics of the action space and environmental factors, which further affects their adaptability in different scenarios.

To address these issues, the paper proposes a comprehensive evaluation framework and benchmark test suite to evaluate the performance of three state - of - the - art VLA models (GPT - 4o, OpenVLA, and JAT) on 20 different datasets. Through this framework, the authors hope to provide tools for the systematic evaluation of future VLA models and identify key improvement directions to promote the development of general - purpose robotic systems.