RoboNav-Pinfo: Personalized Object Navigation Dataset

Expert demonstration trajectories for personalized object-goal navigation in household environments, collected using a Stretch robot in AI2-THOR.

Dataset Summary

	Train	Val	Total
Houses	321	108	429
Trajectories	~40K	~12K	~52K
Total size	~34 GB	~10 GB	~44 GB

Each house contains multiple navigation trajectories (sub-houses), where a robot navigates to find a target object using personalized information (knowledge base text + reference image).

Download & Extract

The dataset is stored as tar archives (one per house) to keep the file count manageable. Download and extract with the following script:

from huggingface_hub import snapshot_download
import tarfile
import os
from pathlib import Path
from tqdm import tqdm

# Download all files
local_dir = snapshot_download(
    repo_id="huluwulu/robonav-pinfo",
    repo_type="dataset",
    local_dir="./robonav-pinfo",
)

# Extract tar archives
for split in ["train", "val"]:
    split_dir = Path(local_dir) / split
    for tar_path in tqdm(sorted(split_dir.glob("*.tar")), desc=f"Extracting {split}"):
        house_id = tar_path.stem  # e.g. "000000"
        out_dir = split_dir / house_id
        out_dir.mkdir(exist_ok=True)
        with tarfile.open(tar_path) as tar:
            tar.extractall(path=out_dir)
        tar_path.unlink()  # remove tar after extraction

After extraction, the directory structure will be:

robonav-pinfo/
├── house_id_to_sub_house_id_train.json
├── house_id_to_sub_house_id_val.json
├── lookup_table.json
├── train/
│   ├── 000000/
│   │   ├── hdf5_sensors.hdf5
│   │   ├── raw_navigation_camera__0.mp4
│   │   ├── raw_manipulation_camera__0.mp4
│   │   ├── pinfo_frame__0.jpg
│   │   ├── ...  (one set per sub-house/trajectory)
│   │   └── success.txt
│   ├── 000001/
│   └── ...
└── val/
    ├── 000321/
    └── ...

File Descriptions

Per-house files

File	Description
hdf5_sensors.hdf5	HDF5 file with trajectory data, keyed by sub-house ID
raw_navigation_camera__{sub_id}.mp4	First-person navigation camera video (224x384)
raw_manipulation_camera__{sub_id}.mp4	Downward-facing manipulation camera video (224x384)
pinfo_frame__{sub_id}.jpg	Reference image of the target object (224x224)
success.txt	Success indicators for each trajectory

HDF5 keys (per sub-house)

Key	Shape	Description
house_index	(T,)	Step indices
templated_task_spec	(T, 928)	Encoded task specification (goal text + KB text)
last_action_str	(T, 200)	Previous action as encoded string
last_action_success	(T, 1)	Whether the previous action succeeded
last_agent_location	(T, 6)	Agent pose [x, y, z, rot_x, rot_y, rot_z]
minimum_l2_target_distance	(T, 1)	L2 distance to target object
hypothetical_task_success	(T, 1)	Whether the task would succeed at this step
nav_task_relevant_object_bbox	group	Bounding boxes of target in nav camera
manip_task_relevant_object_bbox	group	Bounding boxes of target in manip camera

Root-level JSON files

File	Description
house_id_to_sub_house_id_train.json	Maps house ID to list of sub-house IDs (train split)
house_id_to_sub_house_id_val.json	Maps house ID to list of sub-house IDs (val split)
lookup_table.json	Maps between original scene names and numeric house IDs

Usage with RoboNav Training

Point the training script to the extracted dataset directory:

python -m robonav.training.offline.train_pl \
    --model EarlyFusionCnnTransformerPinfo \
    --model_version siglip_base_3 \
    --dataset_version PINFONAV \
    --data_dir ./robonav-pinfo \
    --per_gpu_batch 6 --lr 0.0002 --precision 16-mixed