Structure and usage:
experiments/: experiment scripts for the benchmark. The two main entry points are:benchmark_vlm_as_grounder.py— VLM-as-Grounder experiments (works with any environment)benchmark_vlm_as_planner.py— VLM-as-Planner experiments (works with any environment)benchmark_igibson_oracle.py— Oracle planner baseline for iGibsonvlm_grounder/— helper modules for the VLM-as-Grounder pipeline (state management, checks, planning, VLM queries)
models/: all model classes, most models are implemented with either transformers or vLLM. Extend them to add more modelsplanning/: code for the blocksworld simulated environment and the iGibson clientrendering/: code for the blocksworld renderercode_helpers.py: shared utilities (model loading, domain configuration, logging)
We recommend looking at experiments/ to get started. See the main README for usage examples and experiment variants.
Back to Main Documentation.