All roles with salary

Research Engineer - Robotics RL & Post Training

Microagi MunichEst. Est. EUR 75,000–95,000 / yearMid

Estimated range based on role, country and industry — not published by the company.

Key requirements

  • Python
Accelerating embodied AGI. The next ten years of AI will not be won in software. They will be won in the physical world: in factories, hospitals, kitchens, fields, and homes. The companies that own that data will own the century. microagi is building it. We are a data research and deployment lab. Atlas captures the manual work inside Europe’s industrial champions, trains foundation models on it, and deploys the right robot onto the floor, with no teleoperation and no interruption to the line. Shift records how physical work actually happens and turns it into training-ready data. We were founded by former Formula 1 engineers from Red Bull Racing and Mercedes, together with a researcher from the Alan Turing Institute. Our $55M seed round, the largest in German history, was led by Hummingbird with Northzone, LocalGlobe, Village Global, and redalpine. We are based in Munich, with our research centre in Zürich and teams in London, New York, and Istanbul. We are now building the engineering and research teams that will take Atlas into new industries. What you will work on Train and refine RL policies for contact-rich manipulation, including precise assembly, insertion, and recovery from failures. Develop rewards, curricula, and reset strategies for efficient policy learning. Distill specialist capabilities into generalist policies through expert-generated data, imitation learning, and teacher–student methods. Investigate direct RL post-training of vision-language-action models and world-action models. Build data collection, filtering, sampling, and augmentation pipelines using demonstrations, autonomous rollouts, and human corrections. Build scientifically sound evaluation pipelines across simulation and real robots, measuring success, precision, speed, robustness, and recovery. Design controlled experiments and ablations that isolate the effects of algorithms, data, model changes, and training budgets. Diagnose and reduce sim-to-real gaps in dynamics, sensing, timing, and control. Build reliable, maintainable training and evaluation software with reproducible configurations, versioned data, and traceable results. What we are looking for Strong practical experience implementing and debugging reinforcement learning for continuous control, ideally robotic manipulation. Experience with imitation learning and policy distillation or teacher–student training. Strong working knowledge of ML/DL fundamentals Strong software engineering skills: clear interfaces, maintainable code, meaningful tests, systematic debugging, and reliable experiment infrastructure. Strong Python and PyTorch or JAX skills, with the ability to own work from research implementation through real-robot evaluation. Experience designing rigorous evaluation protocols: meaningful baselines, held-out test conditions, leakage prevention, repeated trials, and appropriate uncertainty estimates. First-principles thinking: the ability to question assumptions, identify underlying mechanisms, and distinguish causal evidence from correlation. Experience with sim-to-real transfer and a robotics simulation stack such as MuJoCo or Isaac Sim / Isaac Lab. Evidence of substantial research contributions through publications at venues such as CoRL, RSS, ICRA, IROS, NeurIPS, ICML, or ICLR, or equivalent demonstrated research and engineering work. Especially relevant experience RL post-training of pretrained VLAs or WAMs. Distilling multiple specialists into a shared generalist while preserving existing capabilities. Visual RL, privileged-information teachers, and contact-rich manipulation. Offline and online RL, human-in-the-loop learning, and learning from failures. Diffusion or flow-matching action policies and temporally chunked actions. Distributed training and large-scale rollout collection. What success looks like Reproducible improvements in real-robot performance, supported by controlled experiments that explain what help

See your match score for this role.

Xecodai maps the interview stages and shows what is preventing a 95% match.

Analyse this role