R2F: Repurposing Ray Frontiers for LLM-free Object Navigation

Francesco Argenziano · John Mark Alexis Marcelo · Michele Brienza · Abdel Hakim Drid · Emanuele Musumeci · Daniele Nardi · Domenico D. Bloisi · Vincenzo Suriani

Setup

conda create -n r2f python=3.9 cmake=3.22 -y && conda activate r2f 
conda install habitat-sim=0.3.0 -c conda-forge -c aihabitathabitat 
pip install -r requirements.txt
python -m spacy download en_core_web_sm   # required for --vln

Model weights for radio_v2.5-b and siglip are downloaded automatically to ckpt/ on first run. HF_HOME is set to ckpt/ automatically via Hydra — to relocate the cache, change hydra.job.env_set.HF_HOME in config/config.yaml.

Data

The evaluation tasks can be found under data/hm3d/val/. Follow the instruction from the official Habitat-Matterport3D repository to download the following scenes: 813, 824, 827, 829, 848, 853, 871, 876, 880, 894. Organize the downloaded data as follows:

data/hm3d/val/
  tasks-objnav.csv     # object-nav tasks
  tasks-vln.csv        # VLN tasks with instruction text
  <scene_hash>/
    <scene>.basis.glb  # 3D scene
    <scene>.glb        # 3D scene with textures
    <scene>.navmesh    # for the path planner
    <scene>.json.gz    # for annotated gt

Running a single task

# Object-nav, task 5, with viewer
python run_tasks.py episodes=5

# Headless
python run_tasks.py episodes=5 no_viewer=true

# VLN mode (NLP instruction parsing is automatic)
python run_tasks.py episodes=5 vln=true no_viewer=true

# Save RGB + similarity heatmap frames
python run_tasks.py episodes=5 dump=true no_viewer=true

Batch run

# All 60 obj-nav tasks, headless
python run_tasks.py episodes=all no_viewer=true

# Subset by range or list
python run_tasks.py episodes=0-9 no_viewer=true
python run_tasks.py 'episodes=0,5,18' no_viewer=true

# Resume an interrupted run
python run_tasks.py episodes=all no_viewer=true resume=true

Key flags

List of all the flags:

Config is in conf/config.yaml and can be overridden from the CLI with key=value syntax.

Key	Default	Description
episodes	required	all, 0-9, or 0,5,18
max_steps	1000	Step budget per episode
map_every	5	Frontier map update interval
no_viewer	false	Headless mode
vln	false	Read tasks-vln.csv, use instruction text as query (NLP parsing automatic)
dump	false	Save RGB + similarity heatmap every 5 steps
resume	false	Skip tasks already in results.csv
seed	42	Random seed

Results are written incrementally to results/<timestamp>/results.csv.

Evaluation

python eval.py                                  # latest results run
python eval.py csv_path=results/<timestamp>/    # specific run
python eval.py json=true                        # machine-readable output

Metrics reported:

Metric	Description
success_rate	Fraction of episodes where agent ended within 1.5m of the target
spl	Success weighted by path length (Anderson et al., 2018)
avg_elapsed_s_on_success	Mean wall-clock time on successful episodes