Open-Ended Instructable Embodied Agents with Memory-Augmented Large Language Models

Gabriel Sarch Yue Wu Michael Tarr Katerina Fragkiadaki

Carnegie Mellon University

EMNLP 2023

Paper

Code

Abstract

Pre-trained and frozen LLMs can effectively map simple scene re-arrangement instructions to programs over a robot's visuomotor functions through appropriate few-shot example prompting. To parse open-domain natural language and adapt to a user's idiosyncratic procedures, not known during prompt engineering time, fixed prompts fall short. In this paper, we introduce HELPER, an embodied agent equipped with as external memory of language-program pairs that parses free-form human-robot dialogue into action programs through retrieval-augmented LLM prompting: relevant memories are retrieved based on the current dialogue, instruction, correction or VLM description, and used as in-context prompt examples for LLM querying. The memory is expanded during deployment to include pairs of user's language and action plans, to assist future inferences and personalize them to the user's language and routines. HELPER sets a new state-of-the-art in the TEACh benchmark in both Execution from Dialog History (EDH) and Trajectory from Dialogue (TfD), with 1.7x improvement over the previous SOTA for TfD.

Memory-Augmented Prompting

A key component of HELPER is its memory of language-program pairs to generate tailored prompts for pretrained LLMs based on the current language context.

The retrieved examples are added to the LLM prompt, which aids in parsing diverse, and user-specific linguistic inputs for planning, re-planning during failures, and interpreting human feedback.

Results

Household Task Execution from Messy Dialogue

We set a new state-of-the-art in the TEACh Trajectory from Dialogue (TfD) and Execution from Dialog History (EDH) benchmarks, where the agent is given a messy dialogue segment and is tasked to infer the sequence of actions from RGB. HELPER improves TfD task success by 1.7x and goal-condition success by 2.1x over existing works with minimal in-domain finetuning.

Task demo

Error correction demo

User Feedback

Gathering user feedback can improve a home robot’s performance, but frequently requesting feedback on a task can diminish the overall user experience. Thus, we enable HELPER to elicit sparse user feedback only when it has completed execution of the program from the initial user input. HELPER improves an additional 1.3X in task success when incorporating just two user feedbacks.

Demo: Clean all cookware with user feedback
(skip to 0:41 for user feedback saying HELPER missed cleaning the pot)

Demo: Make breakfast with user feedback
(skip to 2:54 for user feedback saying did not put tomato & lettuce slice on plate)

User Personalization

HELPER expands its memory of programs with successful executions of user specific procedures; it then recalls and adapts them in future interactions with the user, allowing for user-personalized references.

See our paper for more!

Citation

@inproceedings{sarch2023helper,
                        title = "Open-Ended Instructable Embodied Agents with Memory-Augmented Large Language Models",
                        author = "Sarch, Gabriel and
                        Wu, Yue and
                        Tarr, Michael and
                        Fragkiadaki, Katerina",
                        booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2023",
                        year = "2023"}