Careers · Paris · CDI
Medical Robotics Software Engineer.
Sim2Real · Reinforcement Learning · embedded control. Vendor SDKs ship base locomotion — our proprietary value starts where the vendor stops. You own the Model → Train → Deploy pipeline that gives our robot fleet its care-specific behaviours: safe, reproducible, certifiable.
The role at a glance
Eight facts before you read on.
Why now
The vendor ships the walk. You ship the care.
MedicalCity builds the software infrastructure that makes clinical AI production-ready — an MCP server for medical AI agents, HDS-ready, multi-tenant. Our first deployment combines the agentic platform with a fleet of humanoid and multi-modal robots (LimX TRON 1/TRON 2, humanoids) for in-home and care-home support, funded through France’s public APA benefit.
A signed LOI with the Seine-et-Marne county (CD77) anchors our first pilot, we are raising a €5M seed round, and the CE (MDR) certification path is underway. Vendor SDKs ship base locomotion; the behaviours that matter in care — walking in real homes, gentle manipulation, safe interaction with frail elderly people — are yours to build. This role is the sibling of our Robotics System Engineer: they build the bridge (ROS 2, field deployment); you build the motor brain. You meet on V&V and the CE file.
50-500 Hz
real-time control loop — where your ONNX policies run
100
robots targeted by M12 — every policy you ship multiplies
€5.4B/yr
APA market — solvent, guaranteed by the French state
Your mission
Model. Train. Deploy. Certify.
Model — the simulation twin
Model the platforms: CAD/STL → MJCF/URDF, frames/TF, motor parameters, position-vs-torque control choices. Build MuJoCo environments (plus Isaac Sim/Gazebo) faithful to care settings — homes, care rooms, real obstacles. Calibrate the sim2real gap from real-robot telemetry.
Train — policies & learning
Design Gymnasium environments — states, observations, rewards, terminations, step/reset — with rewards that encode patient safety (capped velocities, minimum distances, controlled falls). Train PPO / GRPO policies with domain randomization for robustness to real homes. Version model weights rigorously: an MDR requirement, not a comfort.
Deploy — from network to robot
Export to ONNX, quantize, run embedded (NVIDIA NX on TRON 1) inside the real-time control loop: action scaling, filtering, software joint limits. Feed policies with IMU, encoders, RealSense D435i, LiDAR. Own latencies, frequencies (50-500 Hz), degraded modes and safe fallbacks.
Evaluate & certify
Define evaluation protocols in simulation and on hardware: fall rate, velocity overshoots, obstacle margins. Run V&V campaigns with the System Engineer — the evidence feeds the CE file (ISO 13482, MDR 2017/745). Apply HDS, GDPR and AI Act requirements to training data and model artefacts.
What we’re looking for
Must-have, and what sets you apart.
Must-have
- Expert Python + PyTorch (or JAX) — you have trained and debugged RL policies yourself, not just followed a tutorial
- Real practice of MuJoCo (MJCF) or Isaac Sim: modelling, physics tuning, Gymnasium pipelines
- At least one successful sim2real transfer on a physical robot — with the pain of the real world: latencies, sensor noise, heating motors
- Solid control foundations: PID, torque control, kinematics/dynamics, filtering
- ONNX (or equivalent) export and embedded inference · Linux, Git, CI/CD
- French at B2+ level (clinical context) · professional English
Strong differentiators
- Legged / humanoid experience (LimX TRON, Unitree Go/G1, ANYmal…) — bipedal locomotion is our daily work
- Whole-body control, MPC, or combining RL with classical control
- VLA / imitation learning exposure — vendor VLA engines and agentic OS are on our product horizon
- Regulated or safety-critical environment — you know an untraceable model is unusable in healthcare
- C++ on the critical paths of the control loop
- Mandarin Chinese — our robot vendors (LimX, Unitree…) and academic ecosystem are Chinese-speaking
- Publications or open-source contributions in robot learning
Honest signal
What this role is not.
Academic research
The goal is a fleet in production, not a paper.
Generalist ML
Here the network outputs motor torques at 50 Hz next to an 87-year-old person.
100 % simulation
You go to the lab and to partner sites to validate your policies on the robots.
A consumer environment
HDS, MDR and AI Act constraints sit at the heart of every technical decision.
Tech stack
The tools of the motor brain.
| Simulation | MuJoCo (MJCF), NVIDIA Isaac Sim, Gazebo, Gymnasium |
| Learning | PyTorch, PPO / GRPO (SB3, RSL-RL, TRL), domain randomization, imitation / BC |
| Deployment | ONNX Runtime, embedded Linux (NVIDIA Jetson/NX), real-time control loop 50-500 Hz |
| Languages | Python (primary), C++ (critical control paths), Bash |
| Robots & sensors | LimX TRON 1 / TRON 2, humanoids · IMU, encoders, RealSense D435i, LiDAR |
| Infra | Docker, Git, CI/CD, experiment tracking (W&B / MLflow), model registry |
| Agentic | MCP (Model Context Protocol) orchestrator interface, LLM orchestration |
What we offer
Policies that matter, next to patients.
- A clear, measurable mission: give the fleet safe, reproducible, certifiable behaviours in real care environments
- Direct impact: your policies run next to patients — and their evidence feeds the CE file
- A diverse fleet: multi-modal bipeds, humanoids, arms — with an open sim2real pipeline you shape from day one
- A tight, senior team: product founders, the System Engineer as your daily counterpart, zero politics
- Compensation: €55-75k gross/year depending on profile + BSPCE (warrants)
- Open stack: you choose your tools, as long as HDS and MDR constraints are met
Interview process — 3-4 weeks, we respond at every step
- 01 Intro call
45 min · video — mission, context, motivations - 02 RL / simulation test
90 min · on-site Paris or video — MJCF modelling, reward design, sim2real debugging - 03 Deep-dive interview
2 h · on-site Paris — with the CTO: pipeline architecture and vision - 04 Sim2real case (optional)
½ day — hands-on on our lab TRON 1: deploy a policy, evaluate it - 05 Comex interview
45 min — cultural and operational alignment - 06 References & offer
Closing
Send CV + a link that beats a cover letter.
A sim2real project, a repo or a video says more than a letter. Reference SWROB-2026-08 · Visa & relocation support available · Also hiring: Robotics System Engineer.