Neurosymbolic World Models for Sequential Decision Making
Authors: Leonardo Hernandez Cano, Maxine Perroni-Scharf, Neil Dhir, Arun Ramamurthy, Armando Solar-Lezama
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the advantages of SWMPO by benchmarking its environment modeling capabilities in a number of simulation tasks. Our experiments aim to answer two research questions: (1) how effectively does SWMPO leverage offline data in the synthesis of an environment-specific FSM?; (2) is the resulting FSM accurate enough for model-based RL? |
| Researcher Affiliation | Collaboration | 1Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute Of Technology, Massachusetts, United States of America 2SIEMENS, New Jersey, United States of America. |
| Pseudocode | Yes | Algorithm 1 Neural Primitives, Algorithm 2 FSMSynth, Algorithm 3 SWMPO, Algorithm 4 greedy Prune |
| Open Source Code | Yes | Implementation details can be found at: https://gitlab. com/da_doomer/swmpo |
| Open Datasets | No | No explicit statement or link is provided for the publicly available datasets used. The paper states that offline data was gathered using controllers in various simulation environments (e.g., 'We use an MPC controller to gather offline data' for Point Mass, 'We use a pre-trained controller provided by the authors to gather offline data' for Li DAR Racing), implying data generation rather than use of pre-existing open datasets. |
| Dataset Splits | No | The paper mentions evaluating models on 'unseen test trajectories' and aggregating errors across 'four test trajectories for each of eight different terrains', but it does not provide specific details on the dataset splits (e.g., percentages for training, validation, and testing, or total sample counts for each split). |
| Hardware Specification | No | The paper does not provide specific details on the hardware used for running experiments, such as GPU or CPU models. |
| Software Dependencies | No | The paper mentions several software packages, such as Pytorch, Scikit-learn, Stable Baselines3, HMM Learn, and SSM, but does not provide specific version numbers for these dependencies. |
| Experiment Setup | Yes | Appendix C provides detailed tables of hyperparameters for each simulation environment, including 'Table 2. Parameters for Point Mass', 'Table 3. Parameters for Autonomous Driving', 'Table 4. Parameters for Salamander', and 'Table 5. Parameters for Bipedal Walker', listing values for parameters such as hidden sizes, learning rate, and batch size. |