Robotics & Controls Portfolio

Clément Dumeril.

I work on model-based control and learning for robots.

MEng student at UC Berkeley (2026-2027), in control of robotic and autonomous systems. Before that, mechanical engineering at Arts et Metiers: CAD, FEA, simulation, machining, vibrations.

See featured work All projects

Academic context

UC Berkeley Mechanical Engineering - Robotics

Arts et Metiers Industrial mechanics

01. Featured Work

Cascaded PID (top) and NMPC (bottom) flying the same figure eight at 5 m/s in gusty wind, flown path in blue over the black reference

Quadrotor Control: C++ NMPC (acados) vs. Cascaded PID in ROS 2

C++ - ROS 2 Jazzy - acados - Webots

A nonlinear MPC in C++ (acados, real-time iteration) flying a DJI Mavic 2 Pro in Webots through ROS 2 at 100 Hz, against a cascaded geometric controller I tuned as well as I could. Without disturbance they are close. In gusty wind the NMPC, with a disturbance observer, halves the tracking error. The repo also lists three bugs that changed my numbers.

0.5 ms median solve at 100 Hz on a 2016 laptop CPU
3.1 vs 7.3 cm RMSE at 5 m/s in gusty wind
1.5 vs 2.3 cm at 5 m/s without disturbance

View code

Paths and error, 5 m/s

RMSE against top speed for each scenario

RMSE vs top speed

ROS 2 graph

NMPC alone

UR5e in Webots grasping a red cube located by the overhead camera

UR5e Neural IK: Differentiable Kinematics + Fine-Tuned Vision

PyTorch - Differentiable kinematics - Webots - Vision (VGG16 fine-tuning)

An MLP that gives the six joint angles of a UR5e for a target position, trained through a PyTorch copy of the arm's forward kinematics. It drives a camera-guided pick-and-place loop in Webots. To find the red cube I fine-tuned a VGG16 (ImageNet weights) on camera images from Webots; a color threshold beat it (8 mm against 49 mm), so the loop uses the threshold. The 12-page preprint reports this and the other things that didn't work.

0.19 mm on held-out targets, 0.53 ms per solve
72x faster than IKPy on the same pose problem
0.33 mm at the grasp point in the closed camera loop

Case study Preprint View code

Type IV hydrogen bottle mission setup with robotic bench

Type IV hydrogen bottle prototype detail on test bench

Type IV hydrogen bottle prototype being carried

Type IV H2 Tank: CalculiX Shells + MLP Correction for a GA Optimizer

CalculiX - NumPy MLP - Mesh convergence - Genetic algorithm (existing)

CalculiX shell model of a Type IV hydrogen vessel on 384 designs, and an MLP that corrects a fast sizing model inside an existing genetic-algorithm optimizer. An audit with 220+ extra runs then showed the dome peak it learned was not mesh-converged. The preprint says so.

384 CalculiX composite-shell designs
Audit: mesh, coverage and end-cap load checks
10-page preprint, single author

Preprint View code

Quiet Ducted Fan: OpenFOAM URANS + FW-H Acoustics, Code_Aster Dynamics

CFD - Modal analysis - Dynamic impact - Technical reporting

Simulation of a quiet desk fan with owl-inspired serrated blades, done at Polytechnia: OpenFOAM CFD and tonal noise for the fan, Code_Aster modes and drop impact for the 3D-printed PETG-CF housing.

43 mN thrust, blade-passing tone at 126 Hz
35-44 Hz first housing mode over the PETG-CF modulus range
Two preprints: aeroacoustics, and structural dynamics

CFD preprint Modes and impact preprint View code

02. Relevant Experience

Graduate Researcher @ UC Berkeley, Agile Robotics and Perception Lab (ARPL)

Advisor: Prof. Giuseppe Loianno. Learning-based control for aerial robotics.

Aug 2026 - Present

Developing learning-accelerated, energy-aware nonlinear MPC for small fixed-wing aerial robots.
Combining learned warm starts, neural approximations and reinforcement learning for real-time trajectory optimization.

Co-founder @ Polytechnia

Engineering consultancy for hardware projects.

May 2026 - August 2026

Won the firm's first client contracts.
Built the fan simulation package (OpenFOAM, Code_Aster) on a 2-node Linux MPI cluster I set up.

Research Assistant @ LCPI

Product design and innovation lab, Arts et Metiers.

September 2025 - June 2026

Fine-tuned a vision-language model to extract engineering data from the sketch and text of industrial idea sheets: ~53% to 76% accuracy.
Compared it with classical computer vision and YOLO.

Commercial & Project Director @ AMJE

Arts et Metiers Junior Etude, Paris.

Mar 2025 - May 2026

Led a 5-engineer team on a $75k Ministry of the Interior R&D mission: a device that stops vehicles without puncturing tires. CAD, simulation, electromechanical and hydraulic integration, 150+ fabrication hours. Alten Prize 2025.
Board member for sales: record $350k annual revenue; trained 30 student project managers.

Project Manager - Engineering Consulting @ AIR FLOW

Part-time mission on railway cooling.

June 2025 - May 2026

Led the team on liquid-nitrogen cooling for a mobile railway setting: thermal studies, heat-exchanger geometry and risk analysis.
Delivered recommendations on safety, performance and compliance.

Engineering Intern @ Groupe Techwave

AI integration at an industrial mechanics and electronics company.

May 2025 - July 2025

Built an LLM assistant (retrieval over engineering documents) for OPEO's consultants.
Contributed to securing EUR 30k in BPI funding through the Accelerateur PME program.

03. All Projects

Everything else, by theme. Pick a card for details.

Robotics & ML

Control, learning and perception.

Simulation

Solvers, FEA and CFD, checked against data.

Software & AI tools

Side builds and applied AI.

Mechanical engineering

Earlier Arts et Metiers projects.

Robotics / Reinforcement Learning / Simulation / Python

Fall 2026

Fixed-Wing: SAC Policy vs. PID in an L2F-Style Simulator

MEng capstone at ARPL (Prof. Loianno): a NumPy fixed-wing simulator on the structure of L2F from RLtools (6-DOF, RK4, gusts, thermals, Dryden turbulence, actuator lag, trim solver) and a Gymnasium task in which a SAC policy (stable-baselines3, 1M steps on Savio) flies the control surfaces. On 100 test flights SAC completes 98 against 100 for a PID, with a lower airspeed error (0.14 vs. 0.37 m/s), but it settles in a 9-degree sideslip that I traced to the reward. A curriculum run then failed because exploration noise alone saturated the cost; the reward now penalizes sideslip and distance to trim, and 3-seed runs with wind are on Savio.

Python / NumPy
SAC (stable-baselines3)
RK4 flight dynamics
Savio (Slurm)

View code on GitHub First SAC run results SAC vs. PID over time

04. Contact

Get in touch.

I'm looking for robotics, controls and ML roles from May 2027. Email is the fastest way to reach me.

GitHub Get in touch Download resume