Research Engineer

ppDo you believe the path to general‑purpose physical AI runs through noisy, real‑world factory deployments? Are you excited by the challenge of turning the classical robotic stacks into the foundational training data for physical AI? Do you want to bridge the gap between world‑class ML research and industrial‑scale robotic execution? /p pIf your answers are yes, we should talk. /p pAt Nomagic, we are executing a humble pivot for general‑purpose physical AI. We believe that physical AI is fundamentally a knowledge transfer problem – we are leveraging the "internet data" of robotics – massive deployment logs from real systems operating in production environments – to bootstrap our efforts. We are looking for Research Engineers who will help us to build, train, and deploy foundational models that bring our fleet from a classical control stack to generalized AI mastery. /p h3Offer essentials /h3 ul liPlay with real robots, solving real problems, every day. /li liRelocation package. /li liFlexible working hours. /li liEnglish‑speaking environment. /li /ul h3What you will do /h3 ul liYour focus will be defined by the intersection of Robotics and ML and large‑scale multimodal model training – expertise in both is optimal and alternatively eagerness to learn. /li liExpect challenges across two main pillars with the opportunity to specialise: ul liCore Research Large‑Scale Infrastructure /li liOwn the Training Stack: Design, implement, and maintain the core infrastructure for large‑scale VLA model training, including scheduling, distribution, job management, checkpointing, and rigorous logging. /li liEnable Rapid Iteration: Build the critical tools and abstractions necessary for launching, monitoring, debugging, and seamlessly reproducing complex, multi‑variant experiments. /li liTrain from Deployment Logs: Utilize our massive repository of offline, classical stack data to pre‑train robust robot foundation models. /li liDrive the Software Feedback Loop: Translate core research needs into concrete infra capabilities, track experiments, analyze results, and close the loop directly with ML researchers to unblock model progress. /li /ul /li liReal‑World Evaluation Operations /li li ul liDesign Physical Benchmarks: Design new robotic tasks and build lightweight physical setups to systematically evaluate model capabilities far beyond the limits of simulation. /li liExecute Structured Evaluations: Ensure robots are properly configured, calibrated, and ready for rollouts. You will coordinate data collection efforts and run structured, on‑robot evaluations to measure real‑world success rates. /li liClose the Physical Feedback Loop: Analyze real‑world evaluation results to guide the ML research direction. You will identify operational bottlenecks across software, hardware, and deployment systems to continuously improve our iteration speed. /li liScale the Workflows: Beta test internal and third‑party tools for teaching robots new skills, and write clear, structured documentation so the broader team can reproduce your workflows and scale your impact. /li /ul /li /ul h3What skills we’d like you to have /h3 ul lipExperience: Deep experience and understanding at the intersection of machine learning, systems engineering, and robotics. /p /li lipProven Track Record: Experience training, fine‑tuning, and deploying modern deep learning architectures (Transformers, VLMs or VLAs, Imitation Learning, RL) for robot control, ideally with policies validated on real hardware. /p /li lipEngineering Excellence: Strong software engineering and infrastructure skills. You are highly proficient in Python and deep learning frameworks (PyTorch/JAX) and can write clean, scalable code for training and evaluation. /p /li lipRobotics Intuition: Comfort working hands‑on with hardware. You understand the robotics full stack (perception, controls, state estimation) and how to debug failures when software meets the physical world. /p /li lipPragmatic Research Mindset: You possess the ability to move seamlessly between research and implementation. You prefer execution, iteration speed, and real‑world robustness over theoretical purity. /p /li /ul /p #J-18808-Ljbffr