Antoine Mahé

Robotics software engineer in a 2-engineer team (plus team lead) building a gantry-based agricultural robot for agro-ecological greenhouses. The system operates like a large-format 3D printer on rails, with a transfer robot that moves it between greenhouse bays. A range of interchangeable tools (seeder, hoe, micro-cultivator, roller, planter, soil-tracking system) sit on a single platform. Stack: Python 3, ROS, SMACH, python-canopen, Git, Docker, Jenkins (CI/CD), Linux, ruff, mypy. • Designed and integrated 7 different farming tools, including drivers for the tool-coupling mechanism (sliding system) using multi-sensor fusion — probes, inductive presence sensors, lock-position sensors — to guarantee safe attachment despite relative motion between greenhouse structure and ground-fixed tool rack. Tuned motor PIDs (variable-frequency drives) per tool via a dedicated tuning utility. • Built a CANopen architecture in Python (python-canopen) implementing the CiA 402 profile (motor control) with inheritance from CiA 301 (sensors, RFID), enabling homogeneous integration of drives from multiple vendors with uneven compliance. Diagnosed and fixed a range of CAN communication issues, from unsuitable sliding contacts for serial bus operation to drive integration mismatches. • Developed SMACH state machines (ROS) for concurrent tasks, interruption, and context-aware resume. Hardened the micro-cultivator from "needs constant supervision" to autonomous operation across a full garden for several hours, through a complete monitoring and error-handling system. On the planter, fine-tuned a controller for a motor driving 3 synchronized mechanisms, with cross-monitored gantry-tool synchronization (mutual abort on error). • Direct work with farmers throughout the project — needs gathering, field testing, agronomic validation. • Deployed on the R&D site in 2022, then on 2 pilot farms (~1 ha each) in 2023. Scaling to 8 robots on a new 4-greenhouse farm (4× previous surface area).

PhD at CentraleSupélec (Metz, LORIA), in collaboration with the DREAM team at Georgia Tech Lorraine and funded by the European INTERREG VA GRoNe project. Thesis: applying machine learning to dynamic system identification in service of model predictive control for mobile robotics. Target controller: MPPI (Model Predictive Path Integral), a variant of MPC well-suited to complex cost functions and non-linear systems, whose effectiveness depends directly on the quality of the dynamic model. • Designed, trained, and benchmarked neural-network dynamic models (MLP, LSTM) as alternatives to classical ARMA/ARX system identification. Comparative benchmark of 300+ architectures (6 architectural families × 4 activation functions) across multiple real and simulated robotic platforms. Stack: Python 3, Keras, TensorFlow, NumPy. • Tackled the imbalanced-data problem inherent to real robotic system identification (near-zero commands dominate the distribution). Developed and compared two training-sample prioritization mechanisms — Prioritized Experience Replay and gradient upper-bound sampling — demonstrating significant performance gains on noisy and imbalanced datasets. • Validated on two physical platforms: a Parrot Bebop2 quadrotor and a Clearpath Kingfisher autonomous boat, in both simulation (Gazebo) and real conditions. On the Kingfisher specifically, integrated the full identification → model → MPPI pipeline onboard an Nvidia Jetson with a Numba-optimized implementation, demonstrating successful autonomous shore-following on an actual body of water. • 4 peer-reviewed publications across ICSTCC 2018, ICAR 2019, and Journal of Intelligent & Robotic Systems (Springer) 2021. See Publications section.

Two missions over four years. • (Feb 2013 – Jun 2013) Internal development: Android (Java) application for diabetes self-monitoring. • (Jun 2013 – Feb 2017) On-site mission at EID (Crédit Mutuel's IT subsidiary), within a team of around ten engineers. Worked on the bank's online card-payment platform — handling roughly 10,000 transactions per hour on a high-availability infrastructure (two sites in active/standby, each redundant). Architecture: payment frontend (internal Python/C framework), C++ backend with MySQL, and integration with the bank's central mainframes responsible for the actual transactions. • PCI-DSS regulated environment: secure code review training, annual third-party audits, strict Git traceability requirements. Full V-model coverage: functional and technical specifications, development, peer security review, qualification by a separate third engineer, then production deployment in coordination with operations (machine-by-machine rollout with verification at each step). Level-3 customer support: production incident diagnosis, log analysis via Bash and Python parsers, root-cause investigation and bug fixes. • Notable project — Monetico Paiement: new card-payment product via mobile terminals, involving external partners (Desjardins bank in Canada, GMX terminal manufacturer). Handled the full V-cycle on the server side, with secure terminal firmware updates as the central technical challenge: implemented asymmetric and symmetric cryptographic flows in C++ via OpenSSL, designed wrappers around X.509 certificate verification and use, and managed terminal-fleet tracking. Significant C++ architecture work (design patterns, component structure). Stack: C++, MySQL, Linux, Git, OpenSSL, STL, Boost, Bash, Python.