Learning with Language-Guided State Abstractions

Authors: Andi Peng, Ilia Sucholutsky, Belinda Z. Li, Theodore Sumers, Thomas L. Griffiths, Jacob Andreas, Julie Shah

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on simulated robotic tasks show that LGA yields state abstractions similar to those designed by humans, but in a fraction of the time, and that these abstractions improve generalization and robustness in the presence of spurious correlations and ambiguous specifications.
Researcher Affiliation Academia Andi Peng MIT Ilia Sucholutsky Princeton Belinda Z. Li MIT Theodore R. Sumers Princeton Thomas L. Griffiths Princeton Jacob Andreas MIT Julie A. Shah MIT
Pseudocode No The paper does not contain any pseudocode or explicitly labeled algorithm blocks.
Open Source Code No The paper does not provide any explicit statement about open-sourcing the code or a link to a code repository.
Open Datasets Yes We generate robotic control tasks from VIMA (Jiang et al., 2022), a vision-based manipulation environment.
Dataset Splits No The paper mentions 'training distributions' and 'test distributions' but does not explicitly describe a validation set or specific train/validation/test splits.
Hardware Specification Yes All computation was done on two NVIDIA Ge Force RTX 3090 GPUs. ... We illustrate the utility of the learned abstractions on mobile manipulation tasks with a Spot robot.
Software Dependencies No The paper mentions software like 'Sentence-BERT' and 'Py Bullet' but does not provide specific version numbers for these or other dependencies.
Experiment Setup Yes We train all networks to convergence for a maximum of 750 epochs.