Robotics Research Engineer (f/m/d)

At NXAI, we are building the future of AI from our hub in Linz, Austria. Our mission is to turn advanced research into real-world impact, with a focus on next-generation xLSTM models and their industrial applications. Working at the intersection of research and industry, we are creating a leading center for AI innovation in the region.

As a Robotics Research Engineer at NXAI, you will take our next-generation sequence models out of the lab and into the physical world. You will work at the intersection of foundation model research and embodied AI, building the perception, reasoning and action pipelines that let robots understand instructions, interpret their environment and act reliably in real world settings.

This is a hands-on research role: you will be as comfortable fine-tuning a Vision-Language-Action (VLA) model as you are debugging a policy on real hardware. You will work closely with our core research team to explore how xLSTM can push robotic foundation models beyond today's state of the art.

Key Responsibilities

• Own physical robot and workstation uptime, including the integration of new sensors, compute units, and end-effectors, while iteratively modifying setups to improve experimental throughput.

• Architect and maintain fault-tolerant, high-throughput pipelines for multi-modal data ingestion, incorporating automated data validation, quality scoring, and versioning.

• Establish standardized, reproducible real-world evaluation protocols and metrics (success rate, task completion time, out-of-distribution robustness) to systematically benchmark policy progression.

• Conduct end-to-end root-cause analysis across the entire robotics stack, diagnosing bottlenecks spanning physical sensors, actuators, network/communication overhead, inference runtimes, and training distribution shift.

• Serve as the critical feedback loop between physical deployment and algorithm design, turning empirical hardware failures and edge cases into actionable research directions and data strategy updates.

• Fine-tune, optimize, and deploy modern learning-based policies (e.g., imitation learning, diffusion policies, vision-language-action models) into real-time, low-latency control loops on physical hardware.

What We Ask

• 2+ years of industry experience training and deploying robot learning models, or an MSc/PhD with research focused on policy learning.

• Demonstrated track record of deploying and debugging neural network policies on physical robots (manipulators, mobile bases, or humanoids), including managing safety constraints and hardware/sensor calibration.

• Advanced proficiency in Python and PyTorch..

• Practical proficiency with Linux environments, modern robotics communication middleware (ROS 1/2), and containerization tools (Docker, Singularity).

• Strong foundational understanding of imitation learning paradigms (e.g., Behavior Cloning, Action Chunking with Transformers [ACT], Diffusion Policy) and dataset curation strategies.

• Discipline in designing reproducible physical benchmarking protocols (success metrics, failure categorization, out-of-distribution testing).

Skills That Will Differentiate You:

• Hands-on work with VLA models or robot foundation models 

• Experience with industrial manipulators or mobile robots in production-like settings

• Sim-to-real transfer and synthetic data generation (e.g. Isaac Sim, MuJoCo) 

• Experience with linear model, such as state-space models, RNNs, or xLSTM

• Publications at robotics or ML venues (ICRA, CoRL, RSS, NeurIPS, ICLR) or open-source contributions

Success Metrics

• Physical setups maintain high operational uptime with minimal experimental downtime. Hardware capabilities (sensing, tooling, compute, and mounting) continuously evolve to directly eliminate data-collection bottlenecks and unlock new research investigations.

• Data collection: a new operator can collect a usable dataset on a robot within a day, with documented setup, synced cameras and joints, and only a small share of episodes discarded for quality issues.

• Evaluation and baselines: a written eval protocol with a fixed number of trials per task (e.g., 20+), plus success rates reported for a baseline policy and a VLA fine-tune, reproducible when rerun.

• Benchmark: a defined set of tasks (e.g., 3–5) with baseline numbers for each, used by the team to compare new methods.

What You’ll Get

• Participation in NXAI’s VSOP program

• Competitive compensation including performance based components

• Opportunity to shape a category defining applied AI company

• Klimaticket

• Hybrid work