Arbitrarily Scalable Environment Generators via Neural Cellular Automata

Authors: Yulun Zhang, Matthew Fontaine, Varun Bhatt, Stefanos Nikolaidis, Jiaoyang Li

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 6 Experimental Evaluation
Researcher Affiliation Academia Yulun Zhang1 Matthew C. Fontaine2 Varun Bhatt2 Stefanos Nikolaidis2 Jiaoyang Li1 1Robotics Institute, Carnegie Mellon University 2Thomas Lord Department of Computer Science, University of Southern California yulunzhang@cmu.edu, {mfontaine,vsbhatt,nikolaid}@usc.edu, jiaoyangli@cmu.edu
Pseudocode No The paper describes methods and processes but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes We include the source code at https://github.com/lunjohnzhang/warehouse_ env_gen_nca_public.
Open Datasets No The paper describes various 'domains' (warehouse, manufacturing, maze) for which environments are generated or used in simulations. While some domains refer to previous works (e.g., [28, 51] for warehouse, [16, 5, 1] for maze), no specific links, DOIs, or explicit statements for public access to the *datasets* used for training or evaluation, or specific author/year citations for these datasets, are provided. For the manufacturing domain, it states, 'The manufacturing domain is new, so we create human-designed environments for it,' but does not provide access.
Dataset Splits No The paper discusses training and evaluation of NCA generators and agent-based simulations (e.g., 'during training' and 'during evaluation' in Section 6, and 'Table 1: Summary of the experiment setup' with S and Seval), but it does not specify any dataset splits like train/validation/test percentages or counts for reproducing data partitioning.
Hardware Specification No Appendix C.3 states, 'This work used Bridge-2 at Pittsburgh Supercomputing Center (PSC) through allocation CIS220115 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program', but it does not provide specific hardware details such as GPU or CPU models, or memory specifications.
Software Dependencies No Appendix C.4 states 'Our implementation uses Python 3.9 and PyTorch [38] with an NVIDIA GPU.' and references 'IBM ILOG CPLEX Optimization Studio [21]', but it does not provide specific version numbers for PyTorch, CPLEX, or other key libraries beyond Python 3.9.
Experiment Setup Yes Table 1 'Summary of the experiment setup' provides specific details such as environment sizes (S=36x33, Seval=101x102), number of agents (Na=200, Na_eval=1400/2250), number of simulations (Ne=5, Neval=10000/100000), NCA iterations (C=50, Ceval=200), and timesteps (T=1000, Teval=5000). Section 6 also discusses hyperparameters like the value in fopt.