Localization · SLAM · Field autonomy
Head of Autonomy at JABAS.AI
Estimation and navigation software for robots that work where GPS doesn't: under canopy, inside polytunnels, between crop rows. The work includes a tightly coupled LiDAR-inertial odometry system built from scratch and a ROS 2 autonomy stack that runs on three robot platforms. It builds on a PhD and first-author papers in RA-L and RAS on long-term 3D LiDAR localization in changing environments.
Focus areas
Factor-graph LiDAR-inertial odometry, IMU preintegration, scan matching, sensor fusion, and long-term localization against maps that change between visits.
Nav2 controller plugins, MPPI, topological route planning, launch orchestration, and safety and metrics tooling for robots in daily commercial use.
Methods go from paper to a robot that runs all day. They are measured on real logs, with benchmarks that report coverage, CPU and failure modes alongside accuracy.
Production work · JABAS.AI navigation stack
JABAS.AI spun out of the University of Lincoln and Ceres Agri-Tech in 2026 to provide autonomy-as-a-service for farm robots. The problem is real and close to home. On soft-fruit farms, up to a fifth of the work is moving trays of picked fruit, often under polytunnel plastic where GPS can't be trusted. Our stack lets robots from different manufacturers do that job without GPS. It runs on the AgileX Hunter, Dynium Gophar and Dogtooth platforms from one codebase, in simulation and on hardware, and is in on-farm trials with commercial growers.
The most rewarding part of the work is the short loop between a derivation on the whiteboard and a robot driving a real polytunnel row, with the field logs showing what to improve next.
Tightly coupled LiDAR-inertial odometry for ROS 2. A sliding-window factor graph fuses 200 Hz IMU with 10 Hz Livox Mid-360 scans and estimates pose, velocity and IMU bias jointly. It was built for crop rows and tunnels, where the geometry is repetitive and scan matching alone degenerates.
C++17 · Python · GTSAM · Eigen · ROS 2 · ~9.6k lines · 12 test suites
A Nav2 controller plugin for Ackermann robots in polytunnel rows. It finds the row boundaries in the costmap or point cloud, infers corridor width on its own, and keeps going when it can only see one side. It travels the exact planned distance even when localization drifts. Validated with a deterministic harness that runs the real Nav2 costmap code: 18 runs and 6,338 control cycles.
C++ · Nav2 plugin API · costmap_2d · GitHub Actions CI
Nav2's latest MPPI controller, backported into the stack with a new four-wheel-steering motion model for the Gophar platform, with tests. It includes "MPPI Control Room", a web UI for live telemetry, velocity monitoring and critic tuning on the robot.
C++ · xtensor · MPPI · JavaScript
Launch and orchestration for the whole stack. It covers GNSS, LIO, point-cloud preprocessing, Nav2, EKF/NavSat fusion, topological navigation, metrics and command safety guards. One codebase handles three platforms on both real hardware and Gazebo, with per-platform Docker devcontainers.
Python launch · Docker · robot_localization · tmule
Contributor to LCAS's open-source topological planner, which has 1,700+ commits. Contributions include the ROS 2 port, per-route-segment metric maps, occupancy-grid publishing, edge actions that set ROS 2 parameters on any node, and topomap editing and image-export tools.
Python · ROS 2 actions · Nav2
Tracks autonomous and manual distance, interventions and collisions, and logs each event to MongoDB with a full system snapshot. It reports mean distance between interventions (MDBI) and is the source of truth for distance-based billing. If MongoDB goes down, the node keeps running and reconnects in the background.
Python · MongoDB · YAML-driven config
Reusable composable nodes for LiDAR preprocessing: transform, crop-box, voxel downsampling, and Patchwork++ ground and outlier removal. They feed both the costmaps and the row controller.
C++ · PCL · rclcpp_components · Patchwork++
A Gazebo Harmonic system plugin that reproduces the Livox Mid-360's non-repetitive scan pattern on a GPU lidar backend. It keeps simulated sensor data close enough to the real sensor for LIO and controller testing.
C++ · Gazebo Sim · ROS 2 Jazzy
Tracks polytunnel poles through occlusion and noisy detections. The tracks give row boundaries for in-row planning and stabilise localization where GPS is unavailable.
C++ · Kalman filtering · data association
Work with JABAS.AI
Want your robot to move autonomously? JABAS.AI brings GPS-free navigation to robots from different manufacturers. Contact the team to talk about your platform.
Research · first author
The PhD research asked which LiDAR points stay stable over weeks and seasons, and how to learn that without manual labels. The result is a family of self-supervised point-stability filters that make scan-to-map localization robust in dynamic environments.
Learns per-point stability from 3D LiDAR alone and generalizes across environments, improving long-term scan-to-map localization. Joint work with Cyrill Stachniss's lab at the University of Bonn.
Unsupervised spatio-temporal labelling plus a stability regression network. Using only long-term stable points gives more reliable localization on long missions than using the full scan.
Compares LIO-SAM, StaticMapping, ORB-SLAM2 and RTAB-Map across four vineyard scenarios with repetitive structure and uneven terrain.
Selected co-authored work
Full list and live citation counts on Google Scholar. Counts on this page as of October 2026.
Experience
2025 – now
Technical lead for localization and navigation at a University of Lincoln and Ceres Agri-Tech spin-out. Work covers the LIO, the controllers and the stack architecture, plus support for on-farm trials with commercial growers.
2024 – 2025
Designed a navigation-agnostic ROS 2 autonomy stack, LiDAR localization for GPS-denied farms, and a Kalman-filter pole tracker for in-row boundary estimation.
2023
Brought robots up on ROS 2: ros2_control, URDF, and sensor and actuator integration for real and simulated platforms.
2020 – 2024
Learning-based long-term localization in dynamic environments. Part of the EU H2020 BACCHUS project.
2018 – 2021
Real-time firmware and Modbus control networks for farm automation. Wrote UART, I²C and SPI drivers, Bluetooth links, over-the-air updates, and fuzzy control.
Education
Contact
For research, talks or questions about LiDAR-inertial odometry, long-term localization and field robotics.
Robot autonomy
Want your robot to drive itself? To make your robot move autonomously, contact the JABAS.AI team.