Getting Started

July 21, 2026 · View on GitHub

Prerequisites

Linux (Ubuntu 20.04+)
NVIDIA GPU, Compute Capability 7.0+, >= 24 GB VRAM
CUDA toolkit (any version supported by your PyTorch)
Conda

1. Install the Evaluator

From source:

git clone <repository_url>
cd <repository_name> && pip install -e .

Verify: python -c "import abotn_evaluator; print('OK')"

2. Deploy the Render Server

The render server provides multi-view RGB observations from 3DGS scenes. It runs in a separate conda environment with PyTorch + CUDA.

# Create env with base dependencies
conda env create -f render_server/environment.yml
conda activate abotn_render

# Install PyTorch matching your CUDA (see https://pytorch.org)
pip install torch --index-url https://download.pytorch.org/whl/cu118  # example for CUDA 11.8

# Compile CUDA extensions
cd render_server
pip install --no-build-isolation ./submodules/diff-plane-rasterization
pip install --no-build-isolation ./submodules/simple-knn

Edit the startup script to set SCENES_ROOT (path to your 3DGS .ply scene files), then start:

bash scripts/start_PointGoal_outdoor_render_server.sh

Script	Task	RENDER_SCALE
`scripts/start_PointGoal_outdoor_render_server.sh`	Point-Goal Outdoor	1.0
`scripts/start_PointGoal_indoor_render_server.sh`	Point-Goal Indoor	1.0
`scripts/start_POIGoal_render_server.sh`	POI-Goal	1.5

Configurable variables in each script:

Variable	Default	Description
`SCENES_ROOT`	—	Path to 3DGS scene directory (required)
`PORT`	7036	HTTP listen port
`GPUS`	`"0"`	GPU IDs, e.g. `"0,1,2"`
`MAX_SCENES_PER_GPU`	1	Scenes loaded per GPU

Ready when logs show Listening on http://localhost:7036/render_gs.

3. Download Data

Dataset	Contents	Download
ABotN-PointBench	31 scenes, 465 trajectories, occupancy maps	🤗 HuggingFace \| 🤖 ModelScope
ABotN-POIBench	11 areas, 163 POIs, occupancy maps	🤗 HuggingFace \| 🤖 ModelScope

Expected layout after extraction:

ABotN-PointBench/
├── Indoor/
│   ├── annotations/          # 16 indoor scenes, each with traj_*.json
│   └── occmaps/              # occupancy maps (same scene structure)
└── Outdoor/
    ├── annotations/          # 15 outdoor scenes, each with traj_*.json
    └── occmaps/

ABotN-POIBench/
├── annotations/              # 11 scenes, each with traj_*.json
└── occmaps/

3DGS scene .ply files should be extracted to a separate directory used as SCENES_ROOT.

When running evaluation, pass annotations/ as --data-dir and occmaps/ as --map-dir.

Prerequisites

1. Install the Evaluator

2. Deploy the Render Server

3. Download Data

Next Steps