Multi-phase Autonomous Small-Body Exploration with Visual SLAM: Bennu and Itokawa Evaluation Dataset

Description and scope This dataset contains the outcome of six simulated closed-loop GNC runs, three around (101955) Bennu and three around (25143) Itokawa. It is intended for testing navigation algorithms for small-body approach and proximity operations. Each run starts about 20 km from the body. An insertion burn 54 h after the start brings the spacecraft to 1-2.5 km from the centre of the body. Our on-board navigation and guidance system, here referred to as the baseline, ran in closed loop, and its estimates drove the pointing and guidance modules. Therefore, control actions, camera pointing, image acquisitions and lidar measurements are fixed to those determined by the baseline. You can run your own navigation algorithm in open loop and compare its errors with those of the baseline. What a sequence contains Ground truth: spacecraft position, velocity and attitude, target attitude and rotation rate, Sun position, executed burns. Sensor data and their models: camera images (2048 × 1536 pixels, 8-bit grayscale PNG, every 60 or 180 s), lidar ranges, star-tracker attitudes. Prior knowledge (perturbed): initial state estimate and covariance, gravity parameter, target rotation, spacecraft mass, solar radiation pressure model. Output of baseline: filter and factor-graph state estimates, landmark map (with truth points added for scoring), commanded burns, metrics. Files and how to start Bennu runs last 95 to 120 h and Itokawa runs 120 h. The six sequences comprise 31,828 images. Two archives hold the sequences: bennu_sequences.zip and itokawa_sequences.zip. common_configs_and_loaders.zip includes shared parameters, metadata, Python and MATLAB loaders, and a ROS 2 bag conversion script. The other files are the README, licence, citation, a sample ROS 2 bag, and a demo video, demo_approach_close_proximity_bennu_itokawa.mp4, showcasing trajectory, images and baseline algorithms execution. The video shows one run around Bennu (top) and one around Itokawa (bottom), from start to end: the trajectory, the camera images with the tracked features, and the map and poses of the SLAM backend. The sample bag, sample_rosbag2_bennu_s0001.zip (188 MB), holds 2 h of bennu_s0001 around the insertion burn. Download every file into one folder and extract common_configs_and_loaders.zip there. The loaders return a sequence's tables, camera intrinsics and baseline products. They need Python 3.12 or later with numpy, pyarrow and pyyaml (and Pillow for the images), or MATLAB R2024b or later with the yaml Add-On by Martin Koch. README.md contains a quick start section with additional information. loaders/ also holds examples in both languages and AGENTS.md, setup instructions for coding agents. loaders/python/convert_to_rosbag2.py writes a sequence as a ROS 2 bag with MCAP storage and also needs the rosbags package. Conventions Tables are Parquet files, and any Parquet reader opens them without the loaders. Tables indexed by time join on the int64 column t_ns, nanoseconds since J2000 in TDB. Column names give quantity, frames and unit, as in p_W_B_m. Quaternions are [w, x, y, z], Hamilton convention. README.md defines the frames, sensor models, simulation, metrics and file layout, and lists the columns, types and units of every table. Licence and citation Data and loaders: CC BY 4.0. Shape models are not included; the target documents give their sources and licences. If you use any of the material of this release, please cite this release (CITATION.cff) and the paper that will be available as part of the conference proceedings: P. Califano, F. Topputo, P. Panicucci. Autonomous Multi-Phase GNC and Visual SLAM for Small-Body Approach, Proximity Operations and Characterization. International Symposium on Artificial Intelligence, Robotics and Automation in Space (i-SAIRAS 2026), Cologne, Germany, 2026. Presented in session TS22 on 4 November 2026. Disclaimer Coding agents have been used to aid the preparation of this dataset. However, none of the content has been generated or uploaded without review. For any inquiry or issue, please contact the corresponding authors.