ML Research Scientist - Time Series Foundation Models 📍 Paris or Lisbon, your choice

About Find & Order

Find & Order is a French–Portuguese deep-tech company with teams in Paris and Lisbon. We build positioning that works where GPS doesn’t: indoors, underground and through outages.

Our technology locates people and equipment indoors without any infrastructure: no beacons, no antennas, no RFID tags. It combines the Earth’s magnetic field, whose indoor distortions act as a unique fingerprint of each building, with the inertial sensors (IMU) already present in phones, wearables and vehicles. Machine learning (Gaussian processes, deep learning), physics and mathematics turn these signals into positions accurate to about 30 cm, on any type of equipment.

Our solutions already help major retail and logistics players such as Carrefour, Auchan, Leroy Merlin, Intermarché and GXO optimise their operations. The company has won multiple awards in France and abroad, including BPI France’s iNOV 2024 competition and the NETVA – Start in America programme.

Our ambition is to push the boundaries of infrastructure-free indoor positioning and become the global technological benchmark.

The role

Our core bet is neural inertial odometry: turning the IMU of a phone, a wearable or a robot into a reliable estimate of how it has moved. Combined with magnetic map matching, it is the motion backbone of our positioning stack.

We are now building an inertial movement foundation model. The central question is whether models pretrained on large, heterogeneous IMU and time-series data can make inertial odometry generalise across devices, body placements and platforms, from pedestrians to robots.

You will own this research end to end: which questions we ask, which models we train, and what reaches the product.

What you’ll do

• Set the research agenda for neural inertial odometry: choose the questions, design experiments with pre-stated kill criteria, and decide what to stop.
• Adapt and build foundation models for IMU data. This means evaluating time-series foundation models (Chronos-style forecasters and their successors) on displacement and trajectory estimation, and fine-tuning or redesigning them where zero-shot use falls short.
• Scale self-supervised pretraining on large multi-device IMU corpora, including multimodal objectives that pair IMU with video or text, then transfer to downstream odometry.
• Grow our data: onboard public pedestrian and robot datasets, and generate synthetic IMU data from motion-capture and motion-generation models.
• Make uncertainty usable, with calibrated covariances and parameter-efficient Bayesian fine-tuning, so that a Kalman filter or factor graph can safely fuse the network’s output.
• Bring geometry into the models: rotation and reflection equivariance, gravity-aligned frames and IMU preintegration.
• Ship and communicate: turn results into models the product can use, and write them up clearly for the team and, where it makes sense, for publication.

What you’ll bring

• Expertise in time-series foundation models. You have trained, fine-tuned or rigorously evaluated models such as Chronos, TimesFM, Moirai, MOMENT or TimeGPT, and you know where they break.
• Hands-on foundation-model experience: large-scale self-supervised or contrastive pretraining, transfer and parameter-efficient fine-tuning (LoRA and relatives) on real GPU budgets.
• Strong deep learning skills in PyTorch, including modern sequence models (Transformers, state-space models such as Mamba, linear attention) and the discipline to debug a training run that quietly goes wrong.
• Genuine interest in inertial odometry for pedestrians, robots or both. Prior IMU work is a plus, not a requirement, if you show you are ready to learn the physics.
• Experimental rigour: baselines, ablations, held-out and out-of-distribution evaluation, and numbers you can trace back to the run that produced them.
• A PhD or equivalent research record in machine learning, robotics, signal processing or a related field. Publications or open-source work are a plus.
• Fluent English. French or Portuguese is a plus, not a requirement.

Nice to have

• Published work on neural inertial odometry or IMU-based motion models (in the line of TLIO, RoNIN, EqNIO, RIO).
• State estimation: EKFs, IMU preintegration, factor graphs (GTSAM or similar), visual-inertial odometry.
• Equivariant or geometric deep learning, and Lie groups as used in robotics.
• Uncertainty quantification: heteroscedastic regression, Laplace approximations, variational or ensemble methods.
• Multimodal pretraining (IMU with video or text), or synthetic sensor data from motion capture and motion generators.
• Experience running a research programme within a small team.

What would really make you stand out

None of these is required, but any one of them will get our attention:

• Gaussian processes: kernels, sparse and scalable inference, GP state-space models. They are already at the heart of our magnetic positioning.
• Neural operators and physics-informed ML: Fourier neural operators, DeepONet, PINNs.
• Bayesian deep learning: calibrated predictive distributions, and telling aleatoric from epistemic uncertainty.
• A passion for physics: you enjoy rigid-body kinematics, sensor models and the maths of motion as much as the models.

Papers close to our research goals

If you have read these and have ideas about where they go next, we want to hear from you:

• EqNIO: Subequivariant Neural Inertial Odometry (Jayanth et al., ICLR 2025)
• Tartan IMU: A Light Foundation Model for Inertial Positioning in Robotics (Zhao et al., CVPR 2025)
• SSIO: Self-Supervised Inertial Odometry (Qiu et al., 2026)
• PRIMUS: Pretraining IMU Encoders with Multimodal Self-Supervision (Das et al., ICASSP 2025)
• Chronos-2: From Univariate to Universal Forecasting (Ansari et al., 2025)

Why join us

• A rare research problem with a direct path to production: your models will run in real buildings for real clients.
• Real ownership of the research agenda within a small, senior team.
• Your choice of office: Paris or Lisbon, with a hybrid work model.
• Competitive salary, continuous training and career development.
• Local benefits depending on location, for example private health and income insurance, meal vouchers and a transport allowance.
• Regular team events across both offices.

General information

• Start date: 1 February 2027
• Long-term contract under local law (French or Portuguese, depending on the office you choose)
• Offices: Av. da República 6, 1050-191 Lisbon, or 2 Rue Henri Chevreau, 75020 Paris
• Education: PhD or equivalent research experience