Research Scientist (Robotics Data & Evaluation) at Innodata

Innodata (Nasdaq: INOD), a global data engineering company, specializes in enabling responsible AI advancement through data, evaluation frameworks, and expertise. Their mission is to empower AI systems trusted at scale by providing solutions for Generative AI and AI builders.

Scope of the Role:

Progress in physical AI is heavily dependent on data design as much as model architecture. The best manipulation and locomotion policies, vision-language-action models, and world models are only as effective as the trajectories, demonstrations, and environments they learn from. Innodata is building the data and evaluation practice behind the next generation of robotics foundation models, partnering directly with customers and frontier labs shaping this frontier. They are hiring a Research Scientist to focus on the science of data for these models.

You will collaborate closely with customers and labs building robot foundation models, focusing on the data these models learn from as much as the models themselves. Your expertise will involve making scientific judgments about data, such as deciding what to capture in the real world versus generating in simulation, how to balance training mixes across different robot datasets, and how to measure whether a policy will perform well outside the lab. Your conclusions will guide what our partners collect next.

What You’ll Own:

• Define how Innodata designs, structures, and evaluates data for robot foundation models, and validate these choices experimentally. This includes:

• Translate requirements of robotics foundation models — such as vision-language-action models, world models, and manipulation and locomotion policies — into concrete data specifications, including modalities, action representations (tokenization, chunking, diffusion and flow-matching action experts), sampling, annotation schemas, and evaluation criteria.

• Decide what is worth capturing in the real world versus generating in simulation, and curate and weight training mixes across heterogeneous robot datasets that span different embodiments, action spaces, and sensor setups.

• Guide collection across capture modalities — motion capture, egocentric, exocentric, teleoperation, and multi-sensor — and across synthetic pipelines, ensuring what is produced aligns with model needs.

• Develop evaluation and benchmarking methodologies that predict real-world transfer, including coverage, discriminative power, reliability, and sim-to-real fidelity. This includes world-model evaluation (rollout quality and action-conditioned prediction) and domain-randomization choices.

• Conduct experiments to validate data quality, fine-tuning and evaluating foundation models on Innodata data, with data-quality ablations and scaling studies to demonstrate specific data decisions improve model performance.

• Design adversarial and long-horizon evaluations to identify where policies and world models fail, and turn these failure modes into improved data.

• Publish findings, contributing benchmarks, methodologies, and papers that advance the field and build trust with partners and labs.

• Collaborate with capture labs, annotation teams, and synthetic-data pipelines to turn specifications into operational collection and labeling plans.

You’ll Thrive in This Role If You Have:

• 4+ years of hands-on industry experience in robot learning or robotics ML. Practical experience is prioritized over formal credentials; a PhD with a compelling research agenda can offset lower experience levels, and senior candidates with the right background are also considered.

• A Bachelor’s degree in computer science, electrical engineering, robotics, or a related technical field is required; an advanced degree (MS or PhD) is preferred.

• Experience training and evaluating robot policies (manipulation or locomotion) using imitation learning or reinforcement learning, with strong PyTorch fundamentals. You should build and measure models, not just call them.

• A strong understanding of datasets: you have curated, filtered, and weighted robot data across different embodiments and sensors, and have clear opinions on what makes data effective for specific objectives.

• Fluency in robotics data formats and standards, including the LeRobot dataset format, RLDS, and Open X-Embodiment, along with common motion and sensor formats.

• Hands-on experience with simulation and synthetic data, including tools like NVIDIA Isaac Sim, Isaac Lab, Omniverse, MuJoCo, and comparable engines. This includes domain randomization, system identification, and sim-to-real transfer.

• Experience with teleoperation or egocentric data collection, and comfort adapting Vision-Language Models (VLM) backbones for control and fine-tuning large Vision-Language-Action (VLA) models using modern toolchains like HuggingFace transformers and PEFT.

• A track record recognized by the field, such as first-author publications or significant open-source contributions at venues like CoRL, ICRA, IROS, RSS, or NeurIPS.

• The ability to work directly with research scientists at partnering labs and explain data and modeling decisions clearly to both expert and non-expert audiences, backed by rigorous, reproducible experiments and documentation.

• Bonus: Interest or hands-on experience in responsible-AI evaluation and red-teaming for embodied systems, such as safety and robustness testing.