Skip to content

Latest commit

 

History

History
15 lines (11 loc) · 1.34 KB

README.md

File metadata and controls

15 lines (11 loc) · 1.34 KB

The modeling baselines are organized into multiple projects (e.g. modeling/vl_model), which are then consumed by the modeling/inference folder for end-to-end mission level evaluation. The models and executors in the modeling/inference/models and modeling/inference/model+executors folders are the end-to-end models and execution strategies, respectively, constructed by using the different modeling projects.

Model Training & Evaluation

1. Placeholder Model

  1. The placeholder robot action prediction model does not require training because it uses a set of heuristic rules to parse the language.
  2. To train and evaluate the vision model (Mask R-CNN) stand-alone for mask generation, please follow the Vision Model README
  3. For evaluating the end-to-end model for mission completion, follow the End-to-End Inference README for the placeholder model.

2. Vision Language Model

  1. To train the VL model for action and mask generation, please follow the Vision Language Model README
  2. To evaluate the VL model for mission completion, follow the End-to-End Inference README for the VL model.

3. Neural-Symbolic Model

  1. To train the Neural-Symbolic model for action prediction, please follow the Neural-Symbolic Model README