Robotics Foundation Models Engineer (human)

Neura Robotics; Freising; On-site; EUR 90,000 - 130,000; Full time

Job summary

Neura Robotics in Germany seeks a senior ML engineer for cognitive robotics, owning production-ready policies from the NeuraGym pipeline. You will work across MAiRA, MiPA, LARA, and humanoid platforms, integrating learned policies with real-time control and safety layers.

You should have 5+ years in robotics ML, experience fine-tuning large models (PEFT/LoRA), and depth in Python, C++, PyTorch, ROS/ROS2, and cloud/simulation tools. Expect frequent travel to customer sites.

Qualifications

• Master's or PhD in Computer Science, Robotics, Electrical Engineering, or a related field.
• 5+ years of hands-on machine-learning experience in robotics, with depth in vision-based manipulation, RL/imitation, or multimodal models.
• Demonstrated experience fine-tuning large pretrained models for robotics; PEFT/LoRA and full fine-tuning, data curation, distillation, and evaluation.

Responsibilities

• Own the outcome on cognitive robotics applications from NeuraGym pipeline to real-world robot deployment.
• Execute the full pipeline: data annotation, training, simulation, deployment, and validation.
• Translate field experience into actionable feedback for product roadmaps and core AI roadmap.

Skills

Python

C++

PyTorch

English (written/spoken)

German (B2-C1)

Problem solving

Education

Master's or PhD in Computer Science / Robotics

Tools

ROS/ROS2

Neura OS

IsaacSim

MuJoCo

TensorFlow

Job description

Your Mission & Challenges

• Own the outcome on the robot: Drive the development of cognitive robotics applications through Physical AI and the NeuraGym platform — you own the outcome on the customer's robot, not just the model in the repo.
• Execute the full pipeline: Take foundation models to production-ready robotic intelligence by executing the complete NeuraGym pipeline — teleoperation, data annotation, training, simulation, deployment, and validation — tailored to real-world robotics applications. Define the data strategy that makes a fine-tune work and own the full post-training loop including hyperparameter iteration, data mixture, LoRA, and honest evaluation.
• Deploy to real-time control: Take policies from checkpoint to real-time closed-loop control — distillation, quantization, action chunking, latency budgets, and on-device/edge inference under the customer's cycle time and reliability requirements.
• Master the full robotics stack: Work across NEURA OS and the control layer — real-time control, kinematics, motion planning, force/impedance control, and the safety architecture — and integrate learned policies cleanly with the low-level controller instead of treating it as a black box.
• Hands-on system integration: Work directly on MAiRA, MiPA, LARA, and humanoid platforms alongside multidisciplinary engineering teams — end-effectors and tool changers, sensor selection and mounting, calibration, electrical and fieldbus interfaces, and the mechanical reality of the cell.
• Enable and support customers: Guide and onboard users to NeuraGym by reviewing training pipelines and scripts, identifying data quality issues, troubleshooting model behavior, and providing hands‑on support to resolve persistent blockers.
• Cross‑functional collaboration: Serve as a key interface between Cloud and AI teams, ensuring smooth collaboration, technical alignment, and efficient problem‑solving across domains.
• Build evaluation rigor: Create the eval harnesses, success‑rate benchmarks, and regression tests that decide whether a policy is ready for handover — and make failure modes visible before the customer finds them.
• Research to production: Prototype techniques from recent papers together with the research team and bring the ones that survive contact with reality into production deployments.
• Close the feedback loop: Translate field experience and customer insights into actionable feedback for both the NeuraGym product roadmap and the NEURA core AI roadmap.

What We Can Look Forward To

• Your education: Master's degree or PhD in Computer Science, Robotics, Electrical Engineering, or a related field.
• ML in robotics depth: 5+ years of hands‑on machine‑learning experience in robotics — strong in at least one of vision‑based manipulation, reinforcement/imitation learning, or multimodal models, and keen to learn the rest.
• Fine‑tuning expertise: Demonstrable experience fine‑tuning large pretrained models for robotics or another domain: PEFT/LoRA and full fine‑tunes, dataset curation, distillation, and evaluation. Familiarity with the open VLA and policy‑learning ecosystem (e.g., π0/π0.5, OpenVLA, LeRobot, diffusion policies) is a strong plus.
• Programming skills: Solid Python skills (C++ is a plus); practical experience with PyTorch or TensorFlow.
• Cloud and simulation: AWS, Azure, or GCP experience and familiarity with robotic simulation tools (IsaacSim, MuJoCo, etc.) are pluses.
• Robotics software stack: Working command of a robotics software stack beyond the model layer — robot operating systems and middleware (NEURA OS, ROS/ROS2 or comparable), real‑time control, kinematics and motion planning, force/impedance control, and how a learned policy wires into a low‑level controller and its safety layer.
• Hardware literacy: Solid understanding of robot hardware — drives and joint modules, torque sensing and encoders, end‑effectors, sensor integration, calibration, electrical and fieldbus interfaces (e.g., EtherCAT), and functional safety concepts — enough to distinguish a model problem from a machine problem on site.
• Execution mindset: Proven problem‑solving abilities, ability to handle multiple projects in parallel, and a strong bias toward execution and personal commitment. You take deployments personally, work through setbacks, and measure yourself by what runs at the customer.
• Communication: Clear communicator who translates between researchers, engineers, and end‑users. You have solid command of English and German (B2–C1 level).
• Travel readiness: Willingness to travel frequently to customer sites and NeuraGyms (approximately 40–60%).