Staff AI Engineer | Robot Autonomy

Cubiq Recruitment; München; Hybrid; EUR 90.000 - 130.000; Vollzeit

Benefits dieser Stelle

ESOP equity

Zusammenfassung

Cubiq Recruitment is seeking a Staff AI Engineer in Munich to own the robot-learning stack across data collection, model training, evaluation and deployment onto physical hardware.

You will teach a physical robot to perform manipulation tasks in real homes, advancing from demonstrations to reliable autonomous behaviour with end-to-end responsibility across engineering and research.

Qualifikationen

• Experience taking a learning-based manipulation system from data collection to deployment on real hardware.
• End-to-end robot learning on real hardware.
• Practical reinforcement learning knowledge.
• Strong Python and modern deep-learning frameworks.
• C++ and strong software-engineering foundations.

Aufgaben

• Design real-world demonstration and data-collection workflows.
• Train manipulation policies using imitation learning and behavioural cloning.
• Apply reinforcement learning where beneficial.
• Evaluate policies on physical robots and analyze failures.
• Deploy models onto real robotic systems.
• Build repeatable training, evaluation and deployment workflows.
• Work with robot platform and data infrastructure engineers.

Kenntnisse

Python

Deep Learning

C++

ROS2

Software Engineering

Robot Learning

Tools

Jetson deployment

NVIDIA GPUs

Diffusion policies

Diffusion transformers

Jobbeschreibung

Teaching a robot to move an object across a table is one problem.

Teaching it to find that object in an unfamiliar home, navigate towards it, pick it up safely and bring it to someone is a very different one.

I’m working with an early-stage humanoid robotics company in Munich that is developing assistive robots for elderly people and their families. The technology has already been tested in real homes, and the next challenge is moving from individual demonstrations towards reliable autonomous behaviour.

They are now hiring a Staff AI Engineer to take ownership of the robot-learning stack across data collection, model training, evaluation and deployment onto physical hardware.

The role

You will teach a physical robot to complete useful manipulation tasks in real domestic environments.

The robot can already perform static pick-and-place tasks and operate light switches. The next behaviours include locating requested objects within a room, navigating towards them, retrieving items from the floor, handing them to a person and helping with tasks such as opening a water bottle.

These tasks are not exceptionally complex in a controlled laboratory. The difficulty is making them work reliably across different homes, objects, lighting conditions, layouts and users.

You will own the complete learning loop rather than focusing on one isolated model or research problem.

What you will be working on

• Designing real-world demonstration and data-collection workflows
• Training manipulation policies using learning from demonstration and behavioural cloning
• Applying reinforcement learning where it provides a practical improvement
• Adapting existing foundation models rather than building a general-purpose model from scratch
• Evaluating policies on physical robots and understanding why they fail
• Deploying models onto real robotic systems
• Improving performance using data collected from deployed robots
• Building repeatable training, evaluation and deployment workflows
• Working closely with the robot platform and data infrastructure engineers
• Helping define the long‑term autonomy architecture as the robot fleet grows

The likely technical direction includes existing VLM and VLA systems, NVIDIA GR00T, Pi-style models, diffusion policies, world models and action-chunking approaches.

The expected balance is approximately 60% engineering and deployment, with 40% research. This is a hands‑on individual‑contributor role rather than a research‑management position.

What they are looking for

• Experience taking a learning-based manipulation system from data collection through to deployment on a physical robot.
• End-to-end robot learning on real hardware
• Manipulation, rather than perception alone
• Learning from demonstration, imitation learning or behavioural cloning
• Practical reinforcement‑learning knowledge
• Strong Python and modern deep‑learning frameworks
• C++ and strong general software‑engineering foundations
• ROS2
• Robot kinematics, dynamics and control
• Training, evaluating and deploying policies on physical robots
• Debugging failures involving calibration, latency, contact, data quality and hardware variation
• Taking ownership of ambiguous technical problems without waiting for a complete specification

A strong publication record is useful, but it will not replace evidence that you have built and deployed a working system.

Candidates whose experience is limited to simulation, object segmentation, grasp detection or isolated motion‑planning research are unlikely to be close enough to the problem.

Useful additional experience

• Diffusion policies or diffusion transformers
• Action chunking
• Bimanual or mobile manipulation
• Humanoid or dexterous robotics
• Jetson deployment and inference optimisation
• Compliant actuation or force-sensitive manipulation
• Continual or distributed learning systems
• Open-source robot-learning projects
• Experience collecting demonstrations through teleoperation

The type of background that could fit

• Humanoid robotics
• Dexterous manipulation
• Warehouse picking
• Academic robot-learning groups with substantial hardware ownership

What’s in it for you?

The product is intended to support elderly people who want to retain more independence within their own homes.

This is not a humanoid programme searching for a future use case. The team has already worked with elderly users and seen genuine demand for the product.

You would be joining early enough to define how the autonomy system is built, while working directly with the CTO in a small and highly technical team.

There is no mature legacy architecture to inherit. The successful hire will help decide the models, tools, data loops and engineering standards that the company builds around.

The work will move quickly from code and experiments to physical robots operating in real environments.

Location and package

• Working model: Predominantly onsite due to the need to work with physical hardware; flexibility to work remotely when the task allows
• Equity: ESOP in addition to salary
• Language: English (Professional Competency)
• Right to work: Candidates cannot receive sponsorship; EU citizens or permanent residency is required
• 20-minute introductory call with the CTO
• Two‑hour technical discussion covering background questions, collaborative architecture of a real problem and a walkthrough of an existing coding project
• Final conversation with the CEO

Apply through the link or message me directly with a short explanation of the robot‑learning system you have personally taken from data collection to real‑hardware deployment.