Long-Horizon Planning for Multi-Agent Robots in Partially Observable Environments

Authors: Sid Nayak, Adelmo Morrison Orozco, Marina Have, Jackson Zhang, Vittal Thirumalai, Darren Chen, Aditya Kapoor, Eric Robinson, Karthik Gopalakrishnan, James Harrison, Anuj Mahajan, Brian Ichter, Hamsa Balakrishnan

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that LLa MAR achieves a 30% higher success rate than other state-of-the-art LM-based multi-agent planners in MAP-THOR and Search & Rescue tasks. Code can be found at https://github.com/nsidn98/LLa MAR
Researcher Affiliation Collaboration 1MIT 2TCS 3USAF-MIT AI Accelerator 4Stanford 5Google Deep Mind 6Apple
Pseudocode Yes The pseudocode for our approach is in Appendix E.
Open Source Code Yes Code can be found at https://github.com/nsidn98/LLa MAR
Open Datasets Yes Additionally, we present MAP-THOR, a comprehensive test suite encompassing household tasks of varying complexity within the AI2-THOR environment. More information about the MAP-THOR and SAR environments can be found in Appendix B and D respectively.
Dataset Splits No The paper specifies training and testing, but does not explicitly mention a validation dataset split or a methodology for it.
Hardware Specification No The paper mentions 'Open AI credits for GPT-4 access' and 'one (1) Apple M1 core' for Sentence BERT fine-tuning, but lacks specific hardware details for the main experiments with LLa MAR.
Software Dependencies Yes We use the clip-vit-large-patch14-336 model for the CLIP weights which we download from https://huggingface.co/openai/clip-vit-large-patch14-336. We finetuned a pre-trained BERT model to function as a semantic mapper between free-form natural language output and the robot s admissible actions in the environment. The pre-trained weights were obtained from https://huggingface.co/sentence-transformers/all-Mini LM-L6-v2.
Experiment Setup Yes Hyperparameters and additional details of the sentence transformer fine-tuning are provided in Appendix J. Epochs 10 Max gradient norm 1 Learning rate 2 10 5 Batch size 64 Encoding dimension 384 Optimizer Adam W Scheduler Warm-up linear Warm-up steps 45 Weight decay 0.01 Loss scale 20 Loss type Multiple negatives ranking loss Similarity function Cosine similarity