Language Models as Zero-Shot Planners: Extracting Actionable Knowledge for Embodied Agents

Authors: Wenlong Huang, Pieter Abbeel, Deepak Pathak, Igor Mordatch

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our evaluation in the recent Virtual Home environment shows that the resulting method substantially improves executability over the LLM baseline. The conducted human evaluation reveals a trade-off between executability and correctness but shows a promising sign towards extracting actionable knowledge from language models.
Researcher Affiliation Collaboration 1University of California, Berkeley 2Carnegie Mellon University 3Google. Correspondence to: Wenlong Huang <wenlong.huang@berkeley.edu>.
Pseudocode Yes Pseudocode is in Appendix A.4. Algorithm 1 Generating Action Plans from Pre-Trained Language Models with Proposed Procedure
Open Source Code Yes Website: https: //huangwl18.github.io/language-planner/.
Open Datasets Yes For our investigation, we use the recently proposed Virtual Home environment (Puig et al., 2018). It can simulate a large variety of realistic human activities in a household environment and supports the ability to perform them via a rich set of 47522 unique embodied actions defined with a verb-object syntax. [...] We use the Activity Programs knowledge base collected by Puig et al. (2018) for evaluation.
Dataset Splits No The paper mentions a "demonstration set" used for prompting and "held-out tasks for evaluation", but does not explicitly define a separate 'validation' split with sizes or percentages for hyperparameter tuning or model selection.
Hardware Specification No The paper does not explicitly provide details about the specific hardware used for running its experiments, such as GPU models, CPU models, or memory specifications.
Software Dependencies No The paper mentions using 'Open AI API', 'Hugging Face Transformers (Wolf et al., 2019)', and 'Sentence Transformers (Reimers & Gurevych, 2019)' but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes For all evaluated methods, we perform hyperparameter search over various sampling parameters, and for methods using a fixed prompt example, we report metrics averaged across three randomly chosen examples. [...] Appendix A.2. Hyperparameter Search