Human's natural language instruction is inherently ambiguous. Standard language grounding and planning methods fail to resolve ambiguity. We propose FISER, which explicitly reasons about human's internal intentions as intermediate steps. 

The robot disamiguates the instruction into a concrete robot-understandable task in the social reasoning phase (Phase 1), and then accomplishes the grounded planning in the embodied reasoning phase (Phase 2). 

Infer Human’s Intentions Before Following Natural Language Instructions

Yanming Wan, Yue Wu, Yiping Wang, Jiayuan Mao*, Natasha Jaques*

Problem Formulation and Method

The following is our problem formulation and proposed method. White nodes represent observable variables, while grey nodes are unobservable. The robot is given the trajectory with a final state at time t', and an utterance u. We propose to explicitly model human's intentions by modeling the human's overall plan G^h as a set of predicates p_k. We further assume that human selects a subgoal p* that needs help, and then specifies a robot's task G^r, which is the underlying intention of human when saying u.

Our FISER framework opts to decompose the problem into two parts -- social reasoning and embodied reasoning.

Specifically, social reasoning is aimed at predicting the sub-task for which the human is asking for assistance, which can be inferred from the context of both the instruction and the observed historical actions of the person in the shared environment. 

After grounding the instructions into robot-understandable tasks, the robot can then do planning and interact with the environment, in a separate embodied reasoning phase. 

To further enhance the model's ability to follow ambiguous instructions, we propose to explicitly add an extra plan recognition stage, where a set of logical predicates is used to help with inferring the human's overall plan.

Transformer-based Model Architecture

We implement a Transformer-based model trained in a supervised learning manner to predict specified sub-tasks (and the human's underlying plan) at intermediate layers. This step-by-step approach distinctly differs from the more commonly employed end-to-end methods in previous works. 

The Transformer-based model has four parts of inputs, which are passed separately into different Transformer Encoder Layers, and interact with each other through a Modality Interaction module after each Transformer layer. The first 2N layers form the social reasoning phase and the last N layers form the embodied reasoning phase. The embeddings at Layer 2N are used for recognizing robot's task, and the last layer embeddings are used for predicting actions.

The HandMeThat Environment

We evaluate our framework by training a Transformer-based model from scratch for the challenging HandMeThat benchmark, and then compare them with multiple competitive baselines, including the state-of-the-art prior work on HandMeThat, and the CoT prompting on the largest available pre-trained language models.

The HandMeThat benchmark introduces a household ambiguous instruction following task rendered in text. Their instructions are split into four difficulty levels, and the gaps between levels correspond to different challenges.

• Level 1: No ambiguity, pure planning. E.g., "give me the book on the sofa."

• Level 2: Social reasoning is required, but it is sufficient to to infer human's goals. E.g., human is storing books into a box and then asks, "could you pass that from the sofa?"

• Level 3: Pragmatic reasoning in language use is further required. E.g., there are multiple books everywhere in the room and only one coat, which is on the sofa. Both a book and a coat are helpful to human's goal, and human asks, "could you pass that from the sofa?" In this case, the human is asking for the book, since "from the sofa" is required to disambiguate which book the human is referring to, but is unnecessary if the human wanted the coat.

• Level 4: Tasks with inherent ambiguities. It cannot be resolved with the existing information, but can potentially be resolved with a strong prior over what human is likely to do. E.g., taking one more fruit will complete the goal of packing picnics, but there are many kinds of fruits in the refrigerator to choose from. From the perspective of completing the goal, any fruit will do, but human preferences may make a difference--the human may want apples instead of bananas at this time.

Results

Key Insights:

1. Explicitly modeling human intentions works better than directly predicting actions.

   • Separating the social and embodied reasoning steps by explicitly recognizing the robot's task is beneficial.

   • Explicitly recognizing the human's plan further helps with the social reasoning stage in the most ambiguous cases (Level 4).

2. Pre-trained LLMs, despite having access to common-sense knowledge, do not adequately perform the complex social and embodied reasoning in this task. Incorporation of domain-specific knowledge through CoT can help.

Conclusion

We study the challenging HandMeThat benchmark, comprising ambiguous instruction following tasks requiring sophisticated embodied and social reasoning. We find that existing approaches for training models end-to-end, or for prompting powerful pre-trained LLMs, are both insufficient to solve these tasks. We hypothesized that performance could be improved by building a model that explicitly performs social reasoning to infer the human's intentions from their prior actions in the environment. Our results provide evidence for this hypothesis, and show that our approach, Follow Instructions with Social and Embodied Reasoning (FISER), enhances performance over the most competitive prompting baselines by 70\%, setting the new state-of-the-art for HandMeThat.

@article{wan2024fiser,

    author    = {Wan, Yanming and Wu, Yue and Wang, Yiping and Mao, Jiayuan and Jaques, Natasha},

    title     = {Infer Human’s Intentions Before Following Natural Language Instructions},

    booktitle = {ArXiv Preprint},

    eprint    = {2409.18073},

    year      = {2024},

}