Differentiable Spatial Planning using Transformers

Authors: Devendra Singh Chaplot, Deepak Pathak, Jitendra Malik

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments to test the effectiveness of the proposed SPT model as compared to prior differentiable planning methods. Datasets. We generate synthetic datasets for training the spatial planning models for both navigation and manipulation settings.
Researcher Affiliation Collaboration 1Facebook AI Research 2Carnegie Mellon University 3UC Berkeley.
Pseudocode No No pseudocode or clearly labeled algorithm block is present.
Open Source Code No Project webpage: https://devendrachaplot.github.io/projects/spatial-planning-transformers. The paper does not explicitly state that source code for the methodology is released at this link or elsewhere, and the link is to a general project page, not a specific code repository.
Open Datasets Yes We generate synthetic datasets for training the spatial planning models for both navigation and manipulation settings. For navigation, we use the Gibson dataset (Xia et al., 2018) to sample maps of size M = 15... License for Gibson dataset: http://svl.stanford.edu/gibson2/assets/GDS_agreement.pdf
Dataset Splits Yes For both the settings, we generate training, validation, and test sets of size 100K/5K/5K maps.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU, GPU models, or memory) used for running experiments.
Software Dependencies No The paper mentions software like the Habitat simulator but does not provide specific version numbers for any key software components or libraries.
Experiment Setup Yes Hyperparameters and Training. For training the SPT model, we use Stochastic Gradient Descent (Bottou, 2010) for optimization with a starting learning rate of 1.0 and a learning rate decay of 0.9 per epoch. We train the model for 40 epochs with a batch size of 20. We use N = 5 Transformer layers each with h = 8 attention heads and a embedding size of d = 64. The inner dimension of the fully connected layers in the transformer is dfc = 512.