Research Scientist / Engineer - Data

Dover; Paris; On-site; EUR 70,000 - 100,000; Full time

Job summary

Dover in Paris hires a Research Scientist / Engineer – Data to develop methods that turn diverse data into better humanoid robot policies. You’ll explore which data improves learning and how to integrate it during pretraining, mid-training, and post-training across on-robot and off-robot data.

Collaborating with AI, Robotics, and Data Infrastructure teams, you will translate ideas into reproducible data processing and training workflows, validating them on humanoid robots and shaping data

Qualifications

• Strong experience in data-centric research, robot learning, computer vision, or multimodal learning.
• Strong Python and PyTorch, with hands-on experience implementing research ideas and training or fine-tuning models.
• A solid understanding of machine learning and experimental design: meaningful baselines, controlled comparisons, reproducibility, and evaluation beyond training loss.
• Experience investigating how data quality, coverage, mixtures, and distribution shifts affect downstream model behavior.
• Experience with multimodal or sequential data, such as video, motion, sensor streams, or action trajectories.
• The ability to connect research findings to practical decisions about data collection, training, and robot evaluation.
• Experience in working with large-scale multimodal data.
• Thrive in a hands-on, fast-paced startup: autonomous, rigorous, collaborative, and curious about robotics and learning.
• Bonus: egocentric data, robot teleoperation, UMI-style demonstrations, simulation, or autonomous-vehicle datasets.
• Bonus: 3D geometry, pose estimation, retargeting, cross-embodiment learning, imitation learning, reinforcement learning, or sim-to-real transfer.
• Bonus: active data selection, public datasets, open-source contributions, or technical writing.
• Bonus: Optimizing data ingestion pipelines during training.

Responsibilities

• Develop methods to make off-robot data useful for robot learning, including action inference, retargeting, and alignment across viewpoints and embodiments.
• Research data composition across training stages: pretraining, mid-training, and post-training using demonstrations, simulation, and deployment experience.
• Design data curation tools and selection methods for filtering, segmentation, annotation, quality scoring, and identifying informative examples.
• Study source mixtures, sampling strategies, curricula, supervision signals, re-annotation and data ingestion strategies, and implement in training experiments.
• Run controlled ablations and scaling studies to isolate effects of data quantity, quality, diversity, and source under realistic compute budgets.
• Measure impact on real-robot success, generalization, robustness, and adaptation efficiency, including how much additional robot data is needed to learn new tasks.
• Shape collection and acquisition priorities based on evidence: which tasks, environments, embodiments, data types, failures, and recoveries are worth capturing, and at what cost.
• Work with data engineers to turn validated research methods into reproducible processing and training workflows.

Skills

data-centric research

robot learning

computer vision

multimodal learning

experimental design

data quality

data mixtures

multimodal data

large-scale data

hands-on startup mindset

egocentric data

cross-embodiment learning

imitation learning

reinforcement learning

sim-to-real transfer

Tools

Python

PyTorch

Job description

Your Mission

As Research Scientist / Engineer – Data, you will develop the methods that efficiently turn diverse data into better humanoid robot policies at UMA. You will investigate which data improves robot learning, how to make it useful, and when to introduce it during training.

You will work with on-robot teleoperation, simulation, and off-robot data—including egocentric human recordings and UMI-style demonstrations—across pretraining, mid-training, and post-training. Your research will help answer central questions: what should we collect next? How can human demonstrations transfer to robots? Which data mixtures, representations, and supervision signals produce the greatest gains in real-world performance?

We already have a working data and training stack and a strong team behind it. Working closely with our AI, Robotics, and Data Infrastructure teams, you will take ideas from dataset design and research prototypes through controlled training experiments to validation on humanoid robots.

Your findings will directly shape our data collection and acquisition strategy, helping us improve robot capabilities while making the best and most efficient use of collection effort and compute.

Key responsibilities:

• Develop methods to make off-robot data useful for robot learning, including action inference, retargeting, and alignment across viewpoints and embodiments, drawing on camera and hand pose estimation or object tracking where useful.

• Research data composition across training stages: broad pretraining, mid-training, and post-training using demonstrations, simulation, and deployment experience.

• Design data curation tools and selection methods for filtering, segmentation, annotation, quality scoring, and identifying informative examples.

• Study source mixtures, sampling strategies, curricula, supervision signals, re-annotation and data ingestion strategies, and implement these ideas in training experiments.

• Run controlled ablations and scaling studies to isolate the effects of data quantity, quality, diversity, and source under realistic compute budgets.

• Measure impact on real-robot success, generalization, robustness, and adaptation efficiency, including how much additional robot data is needed to learn new tasks.

• Shape collection and acquisition priorities based on evidence: which tasks, environments, embodiments, data types, failures, and recoveries are worth capturing, and at what cost.

• Work with data engineers to turn validated research methods into reproducible processing and training workflows.

What You Bring to the Table

• Strong experience in data-centric research, robot learning, computer vision, or multimodal learning, demonstrated through research projects, deployed systems, publications, or open-source work.

• Strong Python and PyTorch, with hands-on experience implementing research ideas and training or fine-tuning models.

• A solid understanding of machine learning and experimental design: meaningful baselines, controlled comparisons, reproducibility, and evaluation beyond training loss.

• Experience investigating how data quality, coverage, mixtures, and distribution shifts affect downstream model behavior.

• Experience with multimodal or sequential data, such as video, motion, sensor streams, or action trajectories.

• The ability to connect research findings to practical decisions about data collection, training, and robot evaluation.

• Experience in working with large-scale multimodal data.

• Thrive in a hands-on, fast-paced startup: autonomous, rigorous, collaborative, and curious about robotics and learning.

• Bonus: egocentric data, robot teleoperation, UMI-style demonstrations, simulation, or autonomous-vehicle datasets.

• Bonus: 3D geometry, pose estimation, retargeting, cross-embodiment learning, imitation learning, reinforcement learning, or sim-to-real transfer.

• Bonus: active data selection, public datasets, open-source contributions, or technical writing.

• Bonus: Optimizing data ingestion pipelines during training.

UMA is an inclusive workplace that values exceptional researchers and builders over perfect resumes. If you enjoy discovering what makes data useful and turning those findings into better robot capabilities, we strongly encourage you to apply even if you don’t tick every box.