Skip to content

User Guide

This guide is organized by the NavArena workflow, covering the full pipeline from raw scenes to evaluation results.

Three-Module Workflow

The three modules execute sequentially, with each step's output feeding the next:

Module Entries

Step Module Description
1 Asset Preprocessing Convert raw 3DGS PLY to V1 unified asset format
2 Data Generator Generate PointNav, VLN, etc. Episode data in preprocessed scenes
3 Evaluation Framework Evaluate navigation models on generated data

Full Workflow Example

# 1. Asset preprocessing (raw PLY -> V1 format)
cd navarena-forge
python -m navarena_forge run-pipeline --config navarena_forge/configs/pipeline.yaml \
    --scene-dir /path/to/raw_scenes/17dc3367 --source-dataset x2robot

# 2. Data generation
cd navarena-gen
python scripts/generate_data.py --config configs/examples/pointnav_example.yaml

# 3. Run evaluation
cd navarena-bench
python -m navarena_bench.scripts.eval --config configs/eval/default_eval.yaml

# 4. Generate replay
python scripts/replay_eval.py --results eval_results/ --output replay.mp4

Common Scenarios

  • Quick single-scene test: Use --num-episodes 5 for generation, --num-episodes 10 for evaluation
  • Batch preprocessing: Use the CLI batch command
  • Remote / ViNT agents: Set agent_type: "remote" or agent_type: "vint" in eval config; see Agents

See also: module documentation linked in the table above, or Core Concepts for architecture and terminology.