Ultra-Fusion: A Resilient Tightly-Coupled Multi-Sensor Fusion SLAM Framework under Sensor Degradation and Spatiotemporal Perturbation for Intelligent Transportation Systems

Core contributors: Yihong Tian, Junjie Zhang, Liuyang Li, and Jie Yin*

Ultra-Fusion is a tightly-coupled multi-sensor SLAM framework for intelligent transportation systems (ITS). It targets real-world deployment where sensor degradation (poor illumination, LiDAR degeneracy, wheel slippage, GNSS outage) and spatiotemporal miscalibration are common.

Within one configurable optimization framework, Ultra-Fusion supports WIO, VIO, LIO, and LVIO, with optional wheel/GNSS fusion and online calibration.

Available releases:

v0.1.0 (ROS1 Noetic) — pre-built binaries to reproduce the paper benchmarks (§2.1–2.2).
v0.1.1 (ROS1 Noetic) — multi-camera support and custom-hardware adaptation; D360 walkthrough (§2.4).
v0.2.0 (ROS2 Humble) — same uf_node + YAML workflow on Ubuntu 22.04 (§2.3).

More demos and releases are on the way — stay tuned.

Overview

Our Goal

We aim to provide a foundational SLAM system for the community: one configurable framework that unifies sensor setups and platforms (wheeled, legged, aerial), and stays ultra-resilient under sensor degradation, timing delays, and extrinsic calibration errors — instead of a pipeline limited to one robot or one sensor stack.

Method Overview

Ultra-Fusion timestamp-orders asynchronous sensor streams and converts them into optional factors inside one sliding-window optimizer, sharing state representation, marginalization, and calibration logic.

Unified pipeline: initialization, reliability scheduling, online calibration, and multi-modal fusion in one framework.

Why Ultra-Fusion

Compared with pipelines tied to a fixed sensor set, Ultra-Fusion focuses on:

Configurability — one framework for WIO / VIO / LIO / LVIO (+ wheel / GNSS).
Reliability — robust localization under corner-case sensor degradation.
Deployability — long-duration and high-speed ITS operation.
Transferability — validated on wheeled, legged, and aerial platforms.
Dual middleware — ROS1 (Noetic) and ROS2 (Humble) runtimes share the same YAML-driven uf_node workflow.

Benchmarks and Findings

Ultra-Fusion is evaluated on five public benchmarks:

M3DGR — wheeled, real + sim, sensor-degradation stress tests
M2DGR-Plus — wheeled, multi-scenario ground SLAM
KAIST Complex Urban — autonomous driving, large-scale urban
GrandTour — legged robots in the wild
MARS-LVIG — aerial LVIG fusion

Reported gains include competitive accuracy and improved localization availability under sensor degradation, calibration perturbation, and long-duration / high-speed operation.

See the project website for videos, benchmark tables, and interactive demos!

Note

This repository currently releases executable binaries and demos. Full source code will be released after paper acceptance. For related implementations, see Ground-Fusion and Ground-Fusion++.

1. Prerequisites & Installation

ROS1 vs ROS2 — which to choose?

	ROS1 Noetic	ROS2 Humble
OS	Ubuntu 20.04	Ubuntu 22.04
Package	v0.1.0 (paper) / v0.1.1 (multi-camera)	v0.2.0
Install	Docker or Native	Same paths — pick the ROS2 image or deps script
Data	ROS1 bags (`rosbag play`)	ROS2 bags (`ros2 bag play`)
Profiles	All five benchmarks + custom device	Any YAML profile — point topics at your ROS2 drivers (M3DGR example)
Docs	§2.1–2.2 · §2.4	§2.3 · ros2 guide

Tip

First time? Start with ROS1 for the full benchmark suite and hardware adaptation guide. Choose ROS2 when your drivers or recordings are already on Humble — Ultra-Fusion is not tied to a single dataset.

The paper release is v0.1.0. v0.1.1 adds multi-camera support and the visual_life reference profile — see §2.4.

Docker installation (recommended) ⭐

Docker images ship only the ROS/runtime stack (middleware, RViz, Ceres, yaml-cpp, system libraries). Install the Ultra-Fusion .deb inside the container after starting it.

Step 1 — Clone this repository (mounted as /workspace in the container)

git clone https://github.com/sjtuyinjie/Ultra-Fusion.git
cd Ultra-Fusion

ROS1 Noetic (Ubuntu 20.04)

Step 2 — Pull the image

# Alibaba Cloud ACR (recommended in China):
docker pull registry.cn-hangzhou.aliyuncs.com/bit_robot_image/ultrafusion:0.1.0

# Docker Hub:
docker pull maotiandocker/ultrafusion:0.1.0

# Or build locally:
docker build -t ultrafusion:0.1.0 .

Step 3 — Start a container (RViz GUI enabled)

Add -v /media:/media:ro if your rosbags live under /media on the host.

xhost +local:docker

docker run --rm -it --net=host --ipc=host \
  -e DISPLAY="${DISPLAY}" \
  -e QT_X11_NO_MITSHM=1 \
  -v /tmp/.X11-unix:/tmp/.X11-unix:rw \
  -v /media:/media:ro \
  -v "$(pwd)":/workspace \
  registry.cn-hangzhou.aliyuncs.com/bit_robot_image/ultrafusion:0.1.0

Step 4 — Install Ultra-Fusion

cd /workspace
./scripts/install_ultrafusion_deb.sh   # add --mirror if GitHub is slow
source /opt/ros/noetic/setup.bash

Step 5 — Verify

which uf_node
rviz -d /opt/ultrafusion/rviz/lio.rviz

Proceed to §2 Running Ultra-Fusion.

ROS2 Humble (Ubuntu 22.04)

Same CMake-based uf_node runtime as ROS1; no colcon build is needed with the prebuilt .deb.

Step	Command
1. Pull image	`docker pull maotiandocker/ultrafusion-ros2:0.2.0` (ACR: `registry.cn-hangzhou.aliyuncs.com/bit_robot_image/ultrafusion-ros2:0.2.0`)
2. Start container	See block below
3. Install `.deb`	`./scripts/install_ultrafusion_ros2_deb.sh`
4. Source ROS2	`source /opt/ros/humble/setup.bash`
5. Run	§2.3

xhost +local:docker

docker run --rm -it --net=host --ipc=host \
  -e DISPLAY="${DISPLAY}" \
  -e QT_X11_NO_MITSHM=1 \
  -v /tmp/.X11-unix:/tmp/.X11-unix:rw \
  -v /media:/media:ro \
  -v "$(pwd)":/workspace \
  maotiandocker/ultrafusion-ros2:0.2.0

cd /workspace
./scripts/install_ultrafusion_ros2_deb.sh
source /opt/ros/humble/setup.bash

Image digests and troubleshooting: docs/ros2_humble_m3dgr.md.

Native installation

For hosts that run Ultra-Fusion directly without Docker.

ROS1 Noetic (Ubuntu 20.04)

Step 1 — Clone and install dependencies

git clone https://github.com/sjtuyinjie/Ultra-Fusion.git
cd Ultra-Fusion
./scripts/install_native_deps.sh

Installs ROS Noetic, PCL/OpenCV/Eigen, and builds Ceres 2.1.0 and yaml-cpp 0.8.0.

Step 2 — Install Ultra-Fusion

./scripts/install_ultrafusion_deb.sh   # add --mirror if GitHub is slow

Step 3 — Source ROS (required in every new shell)

source /opt/ros/noetic/setup.bash

Step 4 — Verify

which uf_node
rviz -d /opt/ultrafusion/rviz/lio.rviz

Proceed to §2 Running Ultra-Fusion.

ROS2 Humble (Ubuntu 22.04)

Dependencies mirror Dockerfile.ros2.

Step 1 — Clone and install dependencies

git clone https://github.com/sjtuyinjie/Ultra-Fusion.git
cd Ultra-Fusion
./scripts/install_native_ros2_deps.sh

Installs ROS2 Humble, PCL/OpenCV/Eigen, rosbags, and builds Ceres 2.1.0 and yaml-cpp 0.8.0.

Step 2 — Install Ultra-Fusion ROS2 package

./scripts/install_ultrafusion_ros2_deb.sh   # add --mirror if GitHub is slow

Step 3 — Source ROS2 (required in every new shell)

source /opt/ros/humble/setup.bash

Step 4 — Verify

which uf_node
rviz2 -d /opt/ultrafusion/rviz/lio_ros2.rviz

Proceed to §2.3 ROS2 runtime.

Installed files

ROS1 package (ultrafusion_*.deb):

Path	Description
`/opt/ultrafusion/bin/uf_node`	Main executable
`/usr/bin/uf_node`, `/usr/bin/uf-node`	CLI shortcuts
`/opt/ultrafusion/config/m3dgr`	M3DGR profiles
`/opt/ultrafusion/config/m2p`	M2DGR-Plus profile
`/opt/ultrafusion/config/lvig`	MARS-LVIG profile
`/opt/ultrafusion/config/kaist`	KAIST profile
`/opt/ultrafusion/config/groundtour`	GrandTour profile
`/opt/ultrafusion/config/visual_life`	Multi-camera LVIO reference profile (D360), v0.1.1
`/opt/ultrafusion/rviz/lio.rviz`	Default RViz layout

ROS2 package (ultrafusion-ros2_*.deb):

Path	Description
`/opt/ultrafusion/bin/uf_node`	Main executable (ROS2 runtime)
`/opt/ultrafusion/config/m3dgr/uf_m3dgr_ros2_*.yaml`	ROS2 YAML profiles
`/opt/ultrafusion/rviz/lio_ros2.rviz`	RViz2 layout

2. Run on Benchmarks & Your Device

Each uf_node <shortcut> maps to a YAML under /opt/ultrafusion/config/. You can also pass a path directly: uf_node /path/to/config.yaml.

Runtime	Source ROS	Play data	Launch	RViz
ROS1	`source /opt/ros/noetic/setup.bash`	`rosbag play … --clock`	`uf_node <shortcut>`	`rviz -d /opt/ultrafusion/rviz/lio.rviz`
ROS2	`source /opt/ros/humble/setup.bash`	`ros2 bag play … --clock`	`uf_node /path/to/config.yaml`	`rviz2 -d /opt/ultrafusion/rviz/lio_ros2.rviz`

ROS1 three-terminal workflow:

Terminal	Command
1	`roscore`
2	`rosbag play /path/to/your.bag --clock`
3	`uf_node <shortcut>`

§2.1–2.2 walk through ROS1 benchmarks; §2.3 covers the ROS2 runtime; §2.4 shows how to adapt your own hardware.

2.1 M3DGR (ROS1)

Wheeled benchmark with real and simulated sensor-degradation sequences. Download rosbags from the M3DGR sequence table, then launch:

# Terminal 1
roscore

# Terminal 2
rosbag play /media/path/to/your.bag --clock

# Terminal 3
uf_node m3dgr                  # default (= m3dgr_standard)
# uf_node m3dgr_standard       # standard LVWIO profile
# uf_node m3dgr_image_enhance  # stronger visual coupling

Released profiles:

Command	Config	Recommended sequences
`uf_node m3dgr`	`.../m3dgr/uf_m3dgr_standard.yaml`	Default M3DGR profile (alias of `m3dgr_standard`)
`uf_node m3dgr_standard`	`.../m3dgr/uf_m3dgr_standard.yaml`	Dynamic01, Varying-illu01, Dark01, Occlusion01
`uf_node m3dgr_image_enhance`	`.../m3dgr/uf_m3dgr_image_enhance.yaml`	Corridor01, GNSS-denial01, Longtime01, Longtime02

Legacy aliases uf_node m3dgr_01 and uf_node m3dgr_02 remain supported.

Expected output (M3DGR) — live LiDAR map + trajectory in RViz:

Corridor01 · uf_node m3dgr_image_enhance
Stable localization in an indoor corridor under vision challenge and LiDAR degeneration. GNSS-denial01 · uf_node m3dgr_image_enhance
Continuous state estimation when GNSS measurements are unavailable. Longtime02 · uf_node m3dgr_image_enhance
Long-duration operation with consistent map and path alignment.

2.2 Other datasets (ROS1)

Additional shortcuts for cross-platform reproducibility. Download rosbags first, then run the matching command. Sequences not listed may need parameter retuning.

Dataset	Platform	Download	demo sequences
M2DGR-Plus	Wheeled ground	Sequence table	Bridge_01, Bridge_02
MARS-LVIG	Aerial UAV	Google Drive rosbags	HKairport01–03, AMtown/AMvalley, HKisland
KAIST Complex Urban	Urban driving	LiDAR + Stereo download	urban25, urban35 (request access on the page)
GrandTour	Legged robot	rosbag download	SPX-2, SNOW-2, EIG-1

Command	Config	Recommended sequences
`uf_node m2p`	`.../m2p/uf_m2p.yaml`	M2DGR-Plus bridge1-style LVWIO
`uf_node lvig`	`.../lvig/uf_lvig.yaml`	MARS-LVIG HKairport01 LVIO
`uf_node kaist`	`.../kaist/uf_kaist.yaml`	KAIST urban25, urban35
`uf_node groundtour`	`.../groundtour/uf_groundtour.yaml`	GrandTour SPX-2, SNOW-2, EIG-1

Expected output (other datasets) — compare your RViz playback with:

GrandTour Arc2 · uf_node groundtour
Legged-platform mapping and trajectory recovery through a large arc-shaped route. KAIST urban25/35 · uf_node kaist
Large-scale LiDAR mapping with high-speed vehicle trajectory visualization. MARS-LVIG HKairport01 · uf_node lvig
Aerial LVIO reconstruction with dense point cloud and flight trajectory.

2.3 ROS2 Humble runtime

v0.2.0 adds a ROS2 Humble runtime alongside ROS1. The workflow is the same: uf_node + YAML profile, ros2 bag play (or live topics), RViz2 for visualization. Any ROS2 dataset works once topics match a profile — copy a released YAML, edit common.* fields, and launch.

Requires ROS2 install (Docker or Native).

Two-terminal workflow:

Terminal	Command
1	`source /opt/ros/humble/setup.bash` → `uf_node /path/to/config.yaml`
2	`ros2 bag play /path/to/your_ros2_bag --clock`

Optional RViz2: rviz2 -d /opt/ultrafusion/rviz/lio_ros2.rviz (fixed frame: world).

Adapt your own ROS2 data:

cp /opt/ultrafusion/config/m3dgr/uf_m3dgr_ros2_lvwio.yaml /tmp/my_ros2.yaml
# edit common.imu_topic, lid_topic, image topics, extrinsics, fusion switches
uf_node /tmp/my_ros2.yaml

Released ROS2 profiles live under /opt/ultrafusion/config/m3dgr/uf_m3dgr_ros2_*.yaml. Field reference: §3 Custom Profiles.

Example: M3DGR on ROS2

M3DGR ships as ROS1 bags. Use the converter script to produce a ROS2 bag with standard common topics, then replay with any ROS2 profile that matches:

python3 -m pip install --user rosbags

python3 scripts/convert_m3dgr_ros1_to_ros2_common.py \
  --src /media/path/to/M3DGR/Grass01.bag \
  --dst /tmp/grass01_20s_ros2 \
  --duration 20 \
  --overwrite

source /opt/ros/humble/setup.bash
uf_node /opt/ultrafusion/config/m3dgr/uf_m3dgr_ros2_lvwio.yaml

source /opt/ros/humble/setup.bash
ros2 bag play /tmp/grass01_20s_ros2 --clock

Verify:

rviz2 -d /opt/ultrafusion/rviz/lio_ros2.rviz
ros2 topic echo /curr_cloud --once --field width
ros2 topic echo /result_lidar_path --once --field header.frame_id

Full-bag conversion, expected topics, and release checklist: docs/ros2_humble_m3dgr.md.

2.4 Adapting Your Own Device (D360 Example)

Copy the closest profile directory, set ROS topics, camera calibration, and extrinsics, then run uf_node with your YAML. Details: docs/visual_life_d360.md (D360 = three fisheye cameras + Livox LiDAR + IMU). YAML fields: §3.

Step	Action
Install	v0.1.1 `.deb` from GitHub Releases
Configure	`cp -a /opt/ultrafusion/config/visual_life /tmp/my_rig` — edit topics, `camera*.yaml`, `multi_camera.modules[]`
Run	`uf_node visual_life` or `uf_node /path/to/config.yaml`
Verify	RViz fixed frame `world`; `/result_path`, `/curr_cloud`, `/feature_reproject_cloud`, `/colored_lidar_cloud`

Multi-camera: use_multi_camera: true and one multi_camera.modules[] entry per stream.

3. Custom Profiles

Released shortcuts are aliases to YAML files under /opt/ultrafusion/config/. To customize, copy the closest profile directory so camera-intrinsic files keep their relative paths. Avoid creating a minimal YAML from scratch — the runtime expects the full field set at startup. Device adaptation walkthrough: §2.4, docs/visual_life_d360.md.

WORK=/tmp/uf_config
mkdir -p "$WORK"
cp -a /opt/ultrafusion/config/m3dgr "$WORK"/

CFG="$WORK/m3dgr/uf_m3dgr_standard.yaml"
${EDITOR:-nano} "$CFG"

roscore &
sleep 3
uf_node "$CFG" &
rosbag play /media/path/to/your.bag --clock

uf_node reads YAML only at startup — restart it after any config change.

3.1 Fusion modes

Keep imu: 1 for the modes below. Visual sensing is selected by use_image. use_gf_standalone_vio is not the UF visual switch: in the current runtime it only selects the Ground-Fusion standalone backend for pure VIO (use_lidar: 0, use_image: 1, wheel: 0). Ultra-Fusion also has its own native VIO/VIWO/LVIO/LVWIO path; keep use_gf_standalone_vio: false for those UF modes.

Target mode	`use_lidar`	`use_image`	`wheel`	`use_gf_standalone_vio`	Runtime path
UF `lvwio`	`1`	`1`	`1`	`false`	Native UF LiDAR + visual + wheel
UF `lvio`	`1`	`1`	`0`	`false`	Native UF LiDAR + visual
UF `vio`	`0`	`1`	`0`	`false`	Native UF standalone VIO (`UFVIO`)
GF standalone `vio`	`0`	`1`	`0`	`true`	Ground-Fusion standalone VIO baseline
UF `viwo`	`0`	`1`	`1`	`false`	Native UF standalone visual + wheel
UF `wio`	`0`	`0`	`1`	ignored	Native UF standalone wheel + IMU
UF `lio`	`1`	`0`	`0`	ignored	LiDAR + IMU
UF `lwio`	`1`	`0`	`1`	ignored	LiDAR + wheel + IMU

Set depth: 1 only for RGB-D visual profiles that really provide the configured depth image; use depth: 0 for monocular RGB visual profiles. use_lidar_reproject only matters for LiDAR+visual profiles, so keep the copied profile's value unless you are intentionally evaluating that coupling. use_planar_wheel_factor selects the planar wheel factor model; keep the released profile's value unless you are intentionally comparing it with the legacy wheel-pose factor.

Sensor topics are configured in common:

common:
  imu_topic: /camera/imu
  lid_topic: /livox/mid360/lidar
  wheel_topic: /odom
  image0_topic: /camera/color/image_raw/compressed
  image1_topic: /camera/aligned_depth_to_color/image_raw

3.2 Camera intrinsics

Camera intrinsics are not stored in the main UF YAML. The main YAML points to camodocal/OpenCV calibration YAML files:

cam0_calib: "color.yaml"
cam1_calib: "color.yaml"

cam0_calib is the primary visual camera file. The runtime loads it as <directory-of-main-config>/<cam0_calib>, so keep the calibration file next to the copied main YAML or preserve the released config directory layout as in the copy example above. In the current runtime this field is treated as a path relative to the main config directory; an absolute path will still be prefixed by that directory.

The released visual profiles use PINHOLE or KANNALA_BRANDT camera models:

%YAML:1.0
---
model_type: PINHOLE
camera_name: camera
image_width: 640
image_height: 480
distortion_parameters:
  k1: 0.0
  k2: 0.0
  p1: 0.0
  p2: 0.0
  k3: 0.0
projection_parameters:
  fx: 607.79772949218
  fy: 607.83526613281
  cx: 328.79772949218
  cy: 245.53321838378

For KANNALA_BRANDT, use projection_parameters fields mu, mv, u0, v0, k2, k3, k4, and k5, following the released fisheye-style calibration files. cam1_calib is only used when the runtime is configured for the two-camera path; for the current single-camera/RGB-D public profiles, keep it consistent with the released template. RGB-D depth input is controlled by depth: 1 and common.image1_topic, not by giving the depth image its own camera-intrinsic YAML.

Multi-camera rigs use multi_camera.modules[].cam_calib instead of cam0_calib — see docs/visual_life_d360.md.

3.3 GNSS fusion

GNSS is independent of the LiDAR/visual/wheel mode switches. UF estimator paths can add raw GNSS pseudorange/Doppler factors and position-only sensor_msgs/NavSatFix factors when the bag provides the required topics. The GF standalone VIO backend receives raw GNSS only; position-only GNSS fixes are not consumed by that backend.

Use case	Main fields	Notes
Disable GNSS	`gnss_enable: 0`	No GNSS subscribers are started
Raw GNSS	`gnss_enable: 1`, `gnss_raw_enable: true`, `gnss_position_enable: false`	Requires range measurements plus ephemeris/iono topics
Position-only GNSS	`gnss_enable: 1`, `gnss_raw_enable: false`, `gnss_position_enable: true`	Uses `sensor_msgs/NavSatFix` in UF estimator paths
Raw + position GNSS	`gnss_enable: 1`, `gnss_raw_enable: true`, `gnss_position_enable: true`	Use only when both measurement types are available

Typical GNSS topic and lever-arm fields:

gnss_meas_topic: /ublox_driver/range_meas
gnss_position_topic: /ublox_driver/receiver_lla
gnss_ephem_topic: /ublox_driver/ephem
gnss_glo_ephem_topic: /ublox_driver/glo_ephem
gnss_iono_params_topic: /ublox_driver/iono_params

gnss_use_antenna_extrinsic: false
gnss_antenna_in_body: [0.0, 0.0, 0.0]

If gnss_use_antenna_extrinsic is true, gnss_antenna_in_body is the antenna position in the estimator body/IMU frame. Do not enable raw GNSS without the matching ephemeris topics; use position-only GNSS in a UF estimator profile or leave GNSS off.

3.4 Extrinsics

All extrinsics are under mapping. Ultra-Fusion uses T_A_B to mean "transform a point from frame B into frame A":

p_A = R_A_B * p_B + t_A_B

Rotation arrays are row-major 3x3 matrices, and translations are in meters.

YAML fields	Transform	Meaning
`extrinsic_T`, `extrinsic_R`	`T_I_L`	LiDAR frame `L` to IMU/body frame `I`
`extrinsic_TIC`, `extrinsic_RIC`	`T_I_C`	Camera frame `C` to IMU/body frame `I`
`extrinsic_TCL`, `extrinsic_RCL`	`T_C_L`	LiDAR frame `L` to camera frame `C`
`extrinsic_TOL`, `extrinsic_ROL`	`T_O_L`	LiDAR frame `L` to wheel/odometer frame `O`
`extrinsic_TIO`, `extrinsic_RIO`	`T_I_O`	Wheel/odometer frame `O` to IMU/body frame `I`

Runtime priority:

Runtime transform	How UF obtains it
`T_I_L`	Always reads `mapping.extrinsic_T/R`
`T_I_C`	Uses `mapping.extrinsic_TIC/RIC` if present; otherwise computes `T_I_L * inverse(T_C_L)` from `extrinsic_TCL/RCL`
`T_C_L`	If `T_I_C` is present, UF also derives internal `T_C_L = inverse(T_I_C) * T_I_L`
`T_I_O`	Uses explicit `mapping.extrinsic_TIO/RIO` if present; otherwise computes `T_I_L * inverse(T_O_L)` from `extrinsic_TOL/ROL`

There is no public YAML flag named estimate_wheel_extrinsic. To change the wheel extrinsic, provide extrinsic_TIO/RIO directly or provide a correct extrinsic_TOL/ROL so UF can derive T_I_O.

3.5 Calibration & delays

For fixed calibration, keep both visual online-calibration flags at zero:

estimate_extrinsic: 0
estimate_td: 0
td: 0.0

In the current runtime, estimate_extrinsic and estimate_td are treated as a joint online camera-IMU calibration request: any nonzero value starts the T_I_C + td calibration state machine after the visual feature and motion excitation gates pass. The state machine first commits T_I_C, then enters the visual delay (td) stage. Therefore estimate_td: 1 alone should not be read as an isolated delay-only mode. This state machine is driven in the UF LiDAR/visual processing path; pure UF standalone VIO/VIWO and GF standalone VIO keep their T_I_C and td parameter blocks fixed in the solver.

Field	Scope	Code behavior
`estimate_extrinsic`	Camera-IMU	`0` does not request online visual calibration by itself; nonzero requests the joint `T_I_C + td` state machine
`estimate_td`	Visual timing	`0` does not request online visual calibration by itself; nonzero also requests the same `T_I_C + td` state machine
`td`	Visual timing	Visual state time uses `image_timestamp + td`
`common.img_time_offset`	ROS image stamp	Added to the ROS image timestamp before visual buffering; this is separate from `td`
`wheel_initial_td`	Wheel timing	Wheel state time uses `wheel_timestamp + wheel_initial_td`
`TimeSync.initial_lidar_to_imu_dt_sec`	LiDAR-IMU timing	Initial LiDAR-to-IMU time offset
`TimeSync.enable_lidar_imu_online_dt`	LiDAR-IMU timing	Enables online LiDAR-IMU time-offset estimation

The current public configs use wheel_initial_td for wheel timing. Legacy fields such as estimate_td_wheel and td_wheel are not the public switch for wheel-delay calibration.

LiDAR-IMU online extrinsic calibration is configured separately:

lidar_imu_calib:
  enable: false
  enable_lock_result: true
  freeze_after_locked_result: true
  apply_locked_result_to_slam: false

This LiDAR-IMU calibrator estimates the rotation part of T_I_L. Locked rotation results affect SLAM only when apply_locked_result_to_slam: true; the LiDAR-IMU translation used by SLAM remains the YAML translation unless you edit the config.

When checking a new profile, inspect the startup log lines for Opti_TIC, Opti_TIO, td, wheel td, GNSS status, and LiDAR-IMU time sync. A smooth but biased trajectory is often a frame or time-offset error, not just solver tuning.

4. Qualitative Results

4.1 Robustness Under Degradation

Representative stress cases: challenging perception conditions with consistent trajectory and map quality.

4.2 Cross-Platform Results

Trajectory estimation examples on ground, legged, and UAV datasets.

For full-scene playback demos, see Expected output in §2.1 and §2.2.

5. License & Acknowledgements

This project is licensed under the MIT License. If you find this project useful, please cite:

@article{tian2026ultra,
  title={Ultra-Fusion: A Resilient Tightly-Coupled Multi-Sensor Fusion SLAM Framework under Sensor Degradation and Spatiotemporal Perturbation for Intelligent Transportation Systems},
  author={Tian, Yihong and Zhang, Junjie and Li, Liuyang and Zhang, Deteng and Zuo, Yunfei and Yin, Jie},
  journal={arXiv preprint arXiv:2606.21223},
  year={2026}
}

Please also consider citing our previous works related to this project:

@article{zhang2025towards,
  author={Zhang, Deteng and Zhang, Junjie and Sun, Yan and Li, Tao and Yin, Hao and Xie, Hongzhao and Yin, Jie},
  booktitle={2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}, 
  title={Towards Robust Sensor-Fusion Ground SLAM: A Comprehensive Benchmark and A Resilient Framework}, 
  year={2025},
  volume={},
  number={},
  pages={8894-8901},
  doi={10.1109/IROS60139.2025.11247507}}


@inproceedings{yin2024ground,
  title={Ground-fusion: A low-cost ground slam system robust to corner cases},
  author={Yin, Jie and Li, Ang and Xi, Wei and Yu, Wenxian and Zou, Danping},
  booktitle={2024 IEEE International Conference on Robotics and Automation (ICRA)},
  pages={8603--8609},
  year={2024},
  organization={IEEE}
}
@article{yin2021m2dgr,
  title={M2dgr: A multi-sensor and multi-scenario slam dataset for ground robots},
  author={Yin, Jie and Li, Ang and Li, Tao and Yu, Wenxian and Zou, Danping},
  journal={IEEE Robotics and Automation Letters},
  volume={7},
  number={2},
  pages={2266--2273},
  year={2021},
  publisher={IEEE}
}

This work is self-funded. Thanks to Tianbao Zhang for providing computation resources. For maintenance, collaboration, or business inquiries, contact maotian616@gmail.com.

Name		Name	Last commit message	Last commit date
Latest commit History 65 Commits
docs		docs
images		images
paper		paper
scripts		scripts
.dockerignore		.dockerignore
.gitignore		.gitignore
Dockerfile		Dockerfile
Dockerfile.ros2		Dockerfile.ros2
LICENSE		LICENSE
README.md		README.md

Navigation Menu

Repository files navigation

Ultra-Fusion: A Resilient Tightly-Coupled Multi-Sensor Fusion SLAM Framework under Sensor Degradation and Spatiotemporal Perturbation for Intelligent Transportation Systems

Contents

Overview

Our Goal

Method Overview

Why Ultra-Fusion

Benchmarks and Findings

1. Prerequisites & Installation

ROS1 vs ROS2 — which to choose?

Docker installation (recommended) ⭐

ROS1 Noetic (Ubuntu 20.04)

ROS2 Humble (Ubuntu 22.04)

Native installation

ROS1 Noetic (Ubuntu 20.04)

ROS2 Humble (Ubuntu 22.04)

Installed files

2. Run on Benchmarks & Your Device

2.1 M3DGR (ROS1)

2.2 Other datasets (ROS1)

2.3 ROS2 Humble runtime

Example: M3DGR on ROS2

2.4 Adapting Your Own Device (D360 Example)

3. Custom Profiles

3.1 Fusion modes

3.2 Camera intrinsics

3.3 GNSS fusion

3.4 Extrinsics

3.5 Calibration & delays

4. Qualitative Results

4.1 Robustness Under Degradation

4.2 Cross-Platform Results

5. License & Acknowledgements

6. Star History

About

Releases

Packages

Contributors

Languages

Folders and files

Latest commit

History

Repository files navigation

Ultra-Fusion: A Resilient Tightly-Coupled Multi-Sensor Fusion SLAM Framework under Sensor Degradation and Spatiotemporal Perturbation for Intelligent Transportation Systems

Contents

Overview

Our Goal

Method Overview

Why Ultra-Fusion

Benchmarks and Findings

1. Prerequisites & Installation

ROS1 vs ROS2 — which to choose?

Docker installation (recommended) ⭐

ROS1 Noetic (Ubuntu 20.04)

ROS2 Humble (Ubuntu 22.04)

Native installation

ROS1 Noetic (Ubuntu 20.04)

ROS2 Humble (Ubuntu 22.04)

Installed files

2. Run on Benchmarks & Your Device

2.1 M3DGR (ROS1)

2.2 Other datasets (ROS1)

2.3 ROS2 Humble runtime

Example: M3DGR on ROS2

2.4 Adapting Your Own Device (D360 Example)

3. Custom Profiles

3.1 Fusion modes

3.2 Camera intrinsics

3.3 GNSS fusion

3.4 Extrinsics

3.5 Calibration & delays

4. Qualitative Results

4.1 Robustness Under Degradation

4.2 Cross-Platform Results

5. License & Acknowledgements

6. Star History

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages