[Paper Review] Towards End-to-End Embodied Decision Making via Multi-modal Large Language Model: Explorations with GPT4-Vision and Beyond

Liang Chen, Yichi Zhang+7 more|arXiv (Cornell University)|Oct 3, 2023

Topic Modeling4 citations

TL;DR

This paper introduces PCA-EVAL, a multi-domain benchmark for evaluating embodied decision-making in perception, cognition, and action, and proposes HOLMES, a multi-agent framework enabling LLMs to leverage multimodal APIs. GPT-4-Vision outperforms GPT-4-HOLMES in end-to-end decision-making with +3% average accuracy, demonstrating state-of-the-art performance, especially in visual understanding, though it remains exclusive to proprietary models.

ABSTRACT

In this study, we explore the potential of Multimodal Large Language Models (MLLMs) in improving embodied decision-making processes for agents. While Large Language Models (LLMs) have been widely used due to their advanced reasoning skills and vast world knowledge, MLLMs like GPT4-Vision offer enhanced visual understanding and reasoning capabilities. We investigate whether state-of-the-art MLLMs can handle embodied decision-making in an end-to-end manner and whether collaborations between LLMs and MLLMs can enhance decision-making. To address these questions, we introduce a new benchmark called PCA-EVAL, which evaluates embodied decision-making from the perspectives of Perception, Cognition, and Action. Additionally, we propose HOLMES, a multi-agent cooperation framework that allows LLMs to leverage MLLMs and APIs to gather multimodal information for informed decision-making. We compare end-to-end embodied decision-making and HOLMES on our benchmark and find that the GPT4-Vision model demonstrates strong end-to-end embodied decision-making abilities, outperforming GPT4-HOLMES in terms of average decision accuracy (+3%). However, this performance is exclusive to the latest GPT4-Vision model, surpassing the open-source state-of-the-art MLLM by 26%. Our results indicate that powerful MLLMs like GPT4-Vision hold promise for decision-making in embodied agents, offering new avenues for MLLM research. Code and data are open at https://github.com/pkunlp-icler/PCA-EVAL/.

Motivation & Objective

• To address the lack of comprehensive benchmarks for evaluating embodied decision-making across perception, cognition, and action in multimodal environments.
• To investigate whether state-of-the-art multimodal large language models (MLLMs) can perform end-to-end embodied decision-making without modality conversion.
• To explore whether collaboration between LLMs and MLLMs via a multi-agent framework can enhance decision-making performance.
• To evaluate the alignment of agent decisions with human values, particularly in safety-critical scenarios.

Proposed method

• Proposes PCA-EVAL, a benchmark with 300 instances across three domains—autonomous driving, domestic assistance, and game-playing—structured as 6-element tuples: image, question, action candidates, answer, reasoning, and key concept.
• Designs a task-specific topology graph to decompose sequential decision-making into one-step problems, enabling fine-grained evaluation beyond final rewards.
• Introduces HOLMES, a multi-agent cooperation framework where an LLM orchestrates multimodal API calls (e.g., object detection, navigation) to gather visual and environmental information.
• Employs GPT-4 and GPT-4-Vision as the core LLM and MLLM, respectively, to compare end-to-end reasoning with API-augmented reasoning in decision-making.
• Uses perception, cognition, and action scores as evaluation metrics, with scores normalized to 1.0 for comparison across models and domains.
• Conducts case studies to analyze modality gap mitigation, decision alignment with human values, and error propagation in multi-step reasoning.

Figure 1: Domain and required ability distribution of PCA-EVAL.

Experimental results

Research questions

• RQ1 Can state-of-the-art multimodal large language models like GPT-4-Vision perform end-to-end embodied decision-making effectively?
• RQ2 How does end-to-end MLLM reasoning compare to LLM-based multi-agent frameworks that use APIs for multimodal perception?
• RQ3 To what extent do MLLMs reduce modality gap and information loss compared to API-mediated LLM approaches?
• RQ4 How well do agent decisions align with human values, especially in safety-critical scenarios like crosswalk navigation?

Key findings

• GPT-4-Vision achieves an average decision accuracy of 0.84 in perception, 0.74 in cognition, and 0.74 in action, outperforming GPT-4-HOLMES by +3% in average accuracy.
• The GPT-4-Vision model improves action score by 26% over the open-source state-of-the-art MLLM, MMICL, demonstrating the performance gap between proprietary and open-source models.
• GPT-4-HOLMES achieves a perception score of 0.88, surpassing GPT-4-Vision’s 0.84, indicating that API-augmented perception can be more accurate in specific visual tasks.
• Despite higher perception accuracy, GPT-4-HOLMES suffers from reasoning error accumulation, resulting in a lower cognition score (0.68) compared to GPT-4-Vision (0.74) in the domestic and game domains.
• Case studies reveal that GPT-4-Vision correctly identifies spatial relationships (e.g., safe driving despite nearby cars), while GPT-4 with APIs misjudges due to missing spatial context.
• The study identifies a critical misalignment: GPT-4 recommends maintaining speed at a crosswalk with no pedestrians, contradicting human safety norms, highlighting the need for value-aligned decision-making.

Figure 2: An instance of PCA-EVAL.