Infer Human's Intentions Before Following Natural Language Instructions

Abstract: For AI agents to be helpful to humans, they should be able to follow natural language instructions to complete everyday cooperative tasks in human environments. However, real human instructions inherently possess ambiguity, because the human speakers assume sufficient prior knowledge about their hidden goals and intentions. Standard language grounding and planning methods fail to address such ambiguities because they do not model human internal goals as additional partially observable factors in the environment. We propose a new framework, Follow Instructions with Social and Embodied Reasoning (FISER), aiming for better natural language instruction following in collaborative embodied tasks. Our framework makes explicit inferences about human goals and intentions as intermediate reasoning steps. We implement a set of Transformer-based models and evaluate them over a challenging benchmark, HandMeThat. We empirically demonstrate that using social reasoning to explicitly infer human intentions before making action plans surpasses purely end-to-end approaches. We also compare our implementation with strong baselines, including Chain of Thought prompting on the largest available pre-trained LLMs, and find that FISER provides better performance on the embodied social reasoning tasks under investigation, reaching the state-of-the-art on HandMeThat.

• The paper introduces the FISER framework, which uses intermediate reasoning to explicitly infer hidden human intentions before executing tasks.
• Transformer-based models using FISER outperform Chain-of-Thought prompting, achieving up to 89% success in clear instruction scenarios.
• The study underscores practical benefits for assistive robotics and domestic automation, highlighting future improvements in scalability and plan recognition.

Infer Human's Intentions Before Following Natural Language Instructions

Overview

The paper "Infer Human's Intentions Before Following Natural Language Instructions" proposes an innovative framework named Follow Instructions with Social and Embodied Reasoning (FISER). The research addresses a critical challenge in enabling AI agents to follow natural language instructions effectively in human-centric environments. Typically, human instructions carry inherent ambiguities, due to assumptions about implicit shared knowledge regarding hidden goals and intentions. The framework aims to enhance AI systems' ability to disambiguate such instructions by explicitly modeling and inferring human goals and intentions as intermediate reasoning steps.

Methodology

The FISER framework is built on the principle that understanding underlying human intentions can lead to better task execution by the AI. The framework is split into two main phases:

1. Social Reasoning: This phase involves human's plan recognition and the robot's task recognition. It aims to infer the sub-task for which human assistance is requested and leverage the observed actions of the person for context.
2. Embodied Reasoning: This phase translates abstract instructions into actionable tasks grounded in the robot's operational scope.

To operationalize this framework, the authors implemented Transformer-based models and evaluated them on the HandMeThat benchmark, which includes multifaceted tasks requiring both social reasoning and physical action planning.

Key Contributions

• Intermediate Reasoning Steps: The introduction of intermediate steps where models explicitly infer human intentions before executing actions. This separation allows for improved disambiguation of natural language instructions.
• Evaluation and Comparison: The research empirically demonstrates that using social reasoning to understand human intentions surpasses purely end-to-end approaches. The performance of models following the FISER framework significantly outstripped several strong baselines, including Chain-of-Thought (CoT) prompting in state-of-the-art LLMs such as GPT-4 Turbo.
• Dataset and Tasks: The study thoroughly tested the models on the HandMeThat (HMT) benchmark, a set of tasks designed to simulate realistic, ambiguous human-robot interaction scenarios.

Results

Transformer-based models trained under the FISER framework showed substantial improvements:

• Level 1: 89.0% success rate, indicating robust planning from clear instructions.
• Level 2: 74.0% success rate, showing the framework's effectiveness in handling ambiguities resolved through context.
• Level 3 and 4: Success rates of ~52.3% and ~51.0% respectively, illustrating efficacy in scenarios requiring deeper inference of human intentions and preferences.

Comparison with LLMs:

• Prompted GPT-4 Turbo, even with Chain-of-Thought and domain-specific enhancements, could not match the efficiency and performance of the bespoke Transformer models developed using the FISER framework. This was evidenced by a marked performance gap, particularly notable in complex, high-ambiguity tasks (Levels 2-4).

Implications and Future Work

The findings underscore the potential of explicitly modeling human intentions in AI systems tasked with interpreting and executing natural language instructions. This approach unlocks more nuanced and contextually aware interactions, critical in domestic and assistive robotics.

Future research could explore:

• Scalability and Generalization: Extending FISER to broader applications and more varied datasets to test adaptability.
• Enhanced Plan Recognition: Improving the precision of human plan recognition to further reduce ambiguities.
• Collaboration with LLMs: Investigating hybrid models that leverage the vast common-sense knowledge of LLMs while adhering to explicit reasoning frameworks like FISER.

Conclusion

The paper presents compelling evidence that the FISER framework is superior in handling instruction ambiguity in collaborative human-robot environments. By decoupling the social reasoning to infer intentions and the embodied reasoning to plan actions, AI systems can achieve more reliable and context-aware performance. This research delineates a significant stride towards creating AI that better understands and integrates into human workflows, enhancing assistive capabilities and domestic labor automation.