Exploratory Retrieval-Augmented Planning For Continual Embodied Instruction Following

This study presents an Exploratory Retrieval-Augmented Planning (ExRAP) framework, designed to tackle continual instruction following tasks of embodied agents in dynamic, non-stationary environments. The framework enhances Large Language Models'(LLMs) embodied reasoning capabilities by efficiently exploring the physical environment and establishing the environmental context memory, thereby effectively grounding the task planning process in time-varying environment contexts. In ExRAP, given multiple continual instruction following tasks, each instruction is decomposed into queries on the environmental context memory and task executions conditioned on the query results. To efficiently handle these multiple tasks that are performed continuously and simultaneously, we implement an exploration-integrated task planning scheme by incorporating the {information-based exploration} into the LLM-based planning process. Combined with memory-augmented query evaluation, this integrated scheme not only allows for a better balance between the validity of the environmental context memory and the load of environment exploration, but also improves overall task performance. Furthermore, we devise a {temporal consistency refinement} scheme for query evaluation to address the inherent decay of knowledge in the memory. Through experiments with VirtualHome, ALFRED, and CARLA, our approach demonstrates robustness against a variety of embodied instruction following scenarios involving different instruction scales and types, and non-stationarity degrees, and it consistently outperforms other state-of-the-art LLM-based task planning approaches in terms of both goal success rate and execution efficiency.

Paper

Similar papers

Peer review

Reviewer kQXj6/10 · confidence 4/52024-06-21

Summary

The paper introduces an Exploratory Retrieval-Augmented Planning (ExRAP) framework, that tackles the problem of embodied planning in dynamic environments. The study focuses on continual instruction following, which is when the task consists of multiple conditional subtasks. The execution of each subtask is dependent on the environment state. This requires continual exploration to keep up-to-date information and integrated planning abilities for efficiency. ExRAP combines Large Language Models (LLMs) with environmental context memory represented as a temporal embodied knowledge graph (TEKG), that captures the states of objects in the scene and their pairwise relationships. The TEKG is updated at each step using the most recent observation and an update function $\mu$. Each subtask is translated into a pair of a query about the scene and a corresponding execution command with an instruction interpreter $\Phi_l$. Next, the memory-augmented query evaluator $\Phi_{M}}$ estimates the likelihood of the query being satisfied, using quadruples retrieved by $\Phi_R$ as environmental context information for the LLM. Authors enforce a temporal-consistency constraint on the evaluator, resulting in a continual steady decrease of the entropy of the answer. The execution commands corresponding to the satisfied queries are then passed to the execution part of the model. This part consists of exploitation and exploration planners. The former assesses the effectiveness of the skill in accomplishing the execution task. The latter one does the same with regard to reducing the uncertainty of the query evaluator $\Phi_M$. The performed skill is chosen as the one providing the maximum weighted sum of the two scores, balancing exploration and exploitation. In the experiments, the authors measure the Success Rate (SR) of the task completion and the Pending Step (PS), representing the average number of steps taken to complete the task. The evaluation is performed on the VirtualHome, ALFRED, and CARLA environments. Performance in various degrees of stationarity in the environment is assessed, showing the advantage of the method. The performance with respect to the number of instructions is also evaluated. The method seems to excel in integrated planning, effectively solving multiple tasks with fewer steps. Ablation studies demonstrate the importance of temporal consistency and exploration-integrated planning. Intrestingly, the study reveals a limited effect of choosing a larger LLM as a base model.

Strengths

- To the best of my knowledge, one of the first works to tackle conditional embodied planning in a dynamically changing environment. - The results of the paper suggest that the method is clearly excels in producing more effective plans than the baselines, integrating multiple tasks in parallel. The plans are being adapted on the fly from the belief about the environment, and the belief is updated by neccessary exploration. This is a valuable contribution.

Weaknesses

- The paper shows that the method is quite robust to the size of the base LLM. It could be a strength if there would be some training procedure possible. However, ExRAP relies exclusively on the in-context learning abilities of the LLM, thus the experiment shows that it scales poorly. The paper would benefit from an ablation study on the size of the retrieval dataset, so it can be shown that the method can scale with the context length of the model by increasing the number of examples. - The method excels at following multiple simple parallel conditional instructions, but it is not clear if it will be as good at the long-term tasks that require multiple steps to achieve, like the ones on which the LLM-Planner was tested. The advantage of having an adaptable planner may become a disadvantage against the baselines, as it may require more queries to the environment context memory, which will create a large computational overhead. It is unclear whether the method is better than the baselines in integrating multiple complex tasks, each of which requires multiple steps, into an optimal plan. Unfortunately, this is not investigated in the paper. - The applicability to the real-world setting is questionable and not investigated. Would the method scale appropriately with the number of objects in the scene? Methods like ConceptGraphs do not show robustness in that case. - The termin "continual embodied instruction following" is a little misleading, as in the paper, the instructions are passed all at once, and not at random times, which would require replanning. The "conditional embodied instruction following in the dynamic environment" could be a better name, though this is not a significant issue.

Questions

- Are the environment settings used in the in-context learning retrieval dataset the same or different from the ones used for the test? - What does the observation $o_t$ consist of for ExRAP? Is it a list of quadruples? How would the setup be adapted for the real-world case? - How does the system handles contradicting instructions? Like “If there is a storm outside, turn off the electricity in the house. If the TV says that the storm is over, turn the electricity on.” - To me, it is not clear why the temporal consistency works. Would a simple baseline of reducing the temperature of prediction perform as good as the temporal consistency constraint in the paper? - It would be interesting to investigate the exploration dynamics of the agent. Does the exploration scores on average decrease steadily throughout the experience or more rapidly in some particular times?

Rating

6

Confidence

4

Soundness

3

Presentation

2

Contribution

2

Limitations

- Runtime overhead not investigated, but mentioned. - Applicability to the real world is not addressed.

Reviewer Jxmz7/10 · confidence 3/52024-07-13

Summary

This paper presents the Exploratory Retrieval-Augmented Planning (ExRAP) framework to address the challenge of continual instruction following in non-stationary embodied environments. ExRAP enhances the reasoning capabilities of Large Language Models (LLMs) by efficiently exploring the environment and maintaining an environmental context memory to ground the task planning process. The paper's main contributions are: 1. A novel ExRAP framework integrating LLMs, memory-augmented reasoning, and information-based exploration for continual instruction following. 2. Temporal consistency refinement and information-based exploration schemes tailored for ExRAP's integrated planning approach.

Strengths

The memory-augmented query evaluation and exploration-integrated planning framework provide a principled way to bridge the gap between high-level language understanding and low-level embodied decision-making. The temporal consistency refinement scheme is a novel contribution in embodied contexts that addresses the important issue of information decay in memory-based solutions over time. Using a temporal embodied knowledge graph (TEKG) as an environmental context memory enables the agent to efficiently assess the satisfaction of instruction conditions without constantly revisiting the environment.

Weaknesses

1. ExRAP's performance still somewhat depends on the underlying LLM's reasoning abilities. While the ablation study shows robustness to smaller LLMs, it's unclear how the framework would perform with much weaker language models or in domains where the LLM's knowledge is limited. 2. The environments' non-stationary aspects are limited to a single instruction changing at fixed intervals. This does not fully capture the complexity and unpredictability of real-world non-stationary environments, where goals, conditions, and constraints can evolve in more intricate ways. 3. The instructions used in the experiments seem to be largely independent of each other, without any significant long-term dependencies or relationships between tasks. An example from Table A.1 is: "If no one is watching the TV, turn it on." If you have an apple somewhere, bring it to your desk. "If you see a book somewhere unorganized, bring it to the sofa." These instructions in the table appear to be independent tasks without any clear long-term dependencies or relationships between them. Each instruction can be completed independently of the others. That being said, the ExRAP framework does take important steps by incorporating memory-augmented reasoning and exploration-integrated planning. These components provide a way for handling evolving instructions and non-stationary environments. The authors' claims may be somewhat overextended given the current experimental setup, but the work still makes valuable contributions to embodied instruction following.

Questions

If authors can address points 2 and 3 from weaknesses, it will help in understanding whether there is a distinction between concurrent or multi-task instruction following and truly continual instruction following.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Authors have addressed some of the limitations of the work in the paper.

Reviewer K6e15/10 · confidence 4/52024-07-14

Summary

The paper considers a problem of continual instruction following where an instruction contains multiple query-execution pairs to be performed. For this, the paper proposes ExRAP comprised of two components: query evaluation and exploration-integrated task planning. Query evaluation incrementally updates the temporal embodied knowledge graph and selects a set of the most relevant executions. The exploration-integrated task planning part uses both exploration and exploitation by weight-summing the predicted values of predefined skills whose weights are hyperparameters. The performance in the experiment shows noticeable margins compared to the baseline models.

Strengths

- The paper is generally written well and easy to follow. - The paper tackles an important and challenging problem of continual instruction following. - Multiple benchmarks are explored for the proposed problem setup, implying generalizability of the setup, yet posing some scale issue (see weaknesses).

Weaknesses

- While agreeing that addressing an instruction that contains multiple tasks, it is unclear why it should be in the explicit form of query(condition)-execution pairs. This is not well-justified. In addition, can't it be implicit? - Regarding the first question, it looks like the proposed approach exploits the query-execution format of the proposed task setup (*e.g.*, query evaluator). - The scale of the continual instruction following dataset is a bit limited. All the 3 datasets use less than 20 continual instructions, which may potentially raise a generalizability concern. - High errors are observed across many tables in the experiments. - Why is this the case? Can this be related to the small scale of the dataset? - One argued contribution is a good performance of the proposed approach, but due to the high errors, it is unclear whether the argued outperformance is valid. - The paper introduces many hyperparameters but it is unclear how much the proposed approach is sensitive to the choice of them. - How to obtain the knowledge graph? is it obtained as GT values or predicted by some learned perception modules? - In L184, why do the authors posit $H(R(q|G_{t-1})) > H(P(q|G_{t-1}))$? Any intuition behind this? - Considering that the term "continual" also refers to learning some new tasks sequentially (*e.g.*, RL [a], IL [b], or even EIF [c]). Discussion about this aspect seems needed to clarify the usage of "continual" in this paper. References\ [a] Xie and Finn. Lifelong robotic reinforcement learning by retaining experiences. CoLLAs, 2022.\ [b] Gao et al. Cril: Continual robot imitation learning via generative and prediction model. IROS, 2021.\ [c] Kim et al. Online continual learning for interactive instruction following agents. ICLR, 2024

Questions

See weaknesses above.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors do not address the potential negative societal impact.

Authorsrebuttal2024-08-11

Thank you for your kind feedback and the additional questions. Our responses are as follows: **For W1.** - Thank you for valuable insight. Our ExRAP can also be applied in situations where instructions consist solely of execution. Below table shows the performance of simple experiments on instructions that involve only execution and no conditions. Although the need for exploration has been reduced, ExRAP demonstrates an advantage over the baseline by efficiently executing multiple instructions through integrated task planning. We will include this results and experiments comparing the entire baselines in the Appendix of the final version. | Model | SR (↑) | PS (↓) | | --- | --- | --- | | LLM-Planner | 74.17% | 8.80 | | ExRAP | 79.86% | 7.20 | - As the reviewer mentioned, in environments where only the robot agent exists and there are no changes over time, the condition of "Always leave the stove open" can be True. However, in dynamic environments where other agents or users can interact with the stove, the instruction would need to be divided into a condition and execution: "If the stove is closed", then "open the stove". For instructions that appear to only involve execution, such as "put the mug in the sink",ExRAP addresses this by relating the condition to the goal, such as “if the mug is not in the sink". If we treat the condition as always true, it could force the repetition of an already completed task, potentially leading to performance degradation. These examples are considered important for explaining our framework, and we will ensure to include them in our paper. - Converting to a query-execution format becomes a crucial part in optimizing exploration in ExRAP. However, simply separating query and execution does not fully resolve the continual instruction following problem in time-varing dynamic environment. We manage environmental observations through an knowledge graph-based approach, and efficiently utilize information through retrieval to evaluate queries. Additionally, an integrated exploration and exploitation planner enables efficient planning through multiple instructions. **For W8.** - Thank you for your thoughtful feedback. We will add the following statement to the main text, Line 108: “The continual instruction does not refer to continual learning, which learns some new task sequentially. It aligns closely with the concept of continual query, which are standing queries that monitor update of interest and return results whenever the update reaches specific thresholds.” We are very pleased that the reviewer's concerns and most of the identified weaknesses have been addressed. If you have any additional questions, please feel free to ask. Thank you!

Reviewer K6e12024-08-12

Thank you for the response.

I thank the authors for the response with the additional experiment. The response addressed my concerns and I am happy to increase the score.

Authorsrebuttal2024-08-12

Thank you!

Thank you for raising your score. We truly appreciate your consideration and constructive comment and feedback. We will incorporate this discussion and additional experiments into the final version. While our paper focus on solving continual instruction, your suggestion to extend our experiments to include what appears to be purely execution-based instruction is insightful feedback that can demonstrate the generality of our model. Thank you once again for your response. The discussion with you is very meaningful and greatly encourages us.

Reviewer K6e12024-08-11

Response by Reviewer K6e1

Thank you for your detailed response. The response addressed my concerns but a few things still remain unclear. - For W1, "Always leave the stove open" seems not implicit but explicit with the condition that is always `True`. The current continual instruction proposed is usually in the form of condition-execution pairs (i.e., if something is happening, then do something). My concern is why we need the "two", condition and execution. For instance, can't it be done with only execution? - For W8, thank you for the clarification, but this can be quite confusing and it might be better to clarify this somewhere in the main paper.

Reviewer Jxmz2024-08-12

Thank you for your detailed response. The response addressed my concerns.

Authorsrebuttal2024-08-12

Thank you!

Thank you once again for your insightful feedback and thoughtful comments. In particular, the discussion on continual instructions with long-term dependencies or relationships between them is invaluable. This results highlight our ExRAP's capability not only to handle independent continual instructions with implicit overlapping subtasks but also to perform effectively in scenarios where these instructions influence each other, emphasizing the generality and scalability of our approach. We will include these additional experimental results and the clarifications requested by the reviewers in the final version of our manuscript. Your review is constructive regarding our approach. Thank you!

Authorsrebuttal2024-08-13

Thank you again for your considerate review. As the end of the author-reviewer discussion period approaches, we summarize the key points of our responses to the reviewer. 1. The scalablity of ExRAP - ExRAP uses the sentence embedding techniques such as DPR and BM25 for knowledge graph retriever and demonstration retriever for exploitation value. Even as the knowledge graph grows, it maintains the inference time of the LLM and minimizes computational overhead through retreival-based approach. By selecting quadruples and demonstrations relevant to queries and instructions, it effectively reduces the context length and maintains the inference time of the LLM. In experiments, we retrieve only 3 demonstrations and 12 quadruples to generate prompts for each query and execution, regardless of the size of the knowledge graph or the dataset. To clarify the retrieval models used, we will add these computational overhead results in Appendix. 2. The applicability to the real-world setting. - The environments used in our experiments (VirtualHome, ALFRED, and CARLA) are highly similar to real-world settings and are frequently used in prior works proposing embodied planning or LLM-based frameworks. For example, in VirtualHome, a single scene contains between approximately 20 to 100 objects, and there are 50 types of room structures, each with varying object states and positions. - In ExRAP, a temporal embodied knowledge graph is used to efficiently manage and retrieve information on changing object states, enhancing adaptability and operational efficiency in dynamic environments. 3. Investigatation for the exploration dynamics of the agent. - Figures 1 in the PDF file illustrate the changes in exploration value for a single query during the execution of continual instructions. We observed that the exploration value increased steadily over time and decreased rapidly once information relevant to the query was collected. Specifically, we noted that the increase in exploration was more larger in environments with higher non-stationarity, leading to enhanced exploration. This, in turn, resulted in a more frequent drop in exploration value. As the reviewer suggested, investigating the exploration dynamics is an interesting analysis that demonstrates how ExRAP improves performance. We will add this to the Appendix of the final version. The reviewer's suggestion to analyze exploration dynamics offers valuable insight by demonstrating the characteristics of our agent's exploration, which helps readers' understanding. Additionally, the discussion on contradicting instructions points to potential approaches for resolving conflicts and offers constructive comments for our future work. We will do our best to address any questions the reviewer may have until the end of the reviewer-author discussion period. Please feel free to ask any further questions. Thank you.

Reviewer kQXj2024-08-13

Thank you for your response. You addressed my concerns completely, and I will raise the score.

Authorsrebuttal2024-08-13

Thank you!

Thank you for raising your score. We greatly appreciate your thoughtful comments and feedback. We will include this discussion and further experiments in our final version. Your recommendation to analyze the exploration dynamics is insightful, as it clarifies how our agent investigates its environment, enhancing the reader's understanding. Furthermore, the discussion on contradicting instructions points to potential approaches for resolving conflicts and offers constructive comments for our future work. Thank you once again for your feedback. Our discussion is meaningful and highly motivating for us.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC