AI Research Engineer

About the job

We're looking for an AI Research Engineer to help build and validate Vision-Language-Action (VLA) and related embodied AI systems for high-dexterity, cross-embodiment robotic manipulation. You will work on model experiments, training and evaluation pipelines, data curation, and sim-to-real validation—partnering with simulation, RL, software, and hardware teams so research outputs become reproducible, deployable assets.

This is a hands-on research engineering role, not a pure theory position. With 1–2 years of experience, you will own well-scoped experiments and infrastructure slices while learning from senior engineers and academic advisors. Success means reliable checkpoints, clear evals, and documented paths from simulation through safe physical testing—not papers in isolation.

What you'll do

• Implement and maintain training and fine-tuning pipelines for VLA and multimodal robot policies (datasets, augmentations, configs, checkpointing, and experiment tracking)
• Curate and validate image-text-action datasets aligned to target embodiments, tasks, and hardware constraints; document splits, metadata, and quality checks
• Design and run evaluation harnesses in simulation (e.g. Isaac Sim, MuJoCo) and on physical robots—success metrics, regression suites, latency and failure-mode logging
• Support cross-embodiment experimentation: adapting architectures, action spaces, and training logic so models generalize across robot morphologies under team guidance
• Contribute to data and deployment infrastructure—monitoring hooks, inference workflows (Docker / ROS-style), and handoffs that keep fleet-level reliability in view
• Run structured sim-to-real experiments: domain randomization studies, ablations, and documented protocols for bridging synthetic training to hardware behavior
• Analyze failures (occlusion, drift, hallucinated actions, contact issues) and turn findings into actionable recommendations for data collection, model changes, or sim improvements
• Participate in technical reviews and roadmap syncs with engineering leads; incorporate advisor and teammate feedback into production-ready deliverables
• Maintain clean, documented code, runbooks, and experiment reports so others can reproduce and extend your work

What we're looking for

• 1–2 years of experience in ML research engineering, robotics ML, or a closely related internship plus early full-time role (or equivalent strong project portfolio)
• Degree in Computer Science, Robotics, AI, Electrical Engineering, or related field—or demonstrable equivalent through shipped experiments and open or internal repos
• Working proficiency in Python and PyTorch (JAX or similar acceptable); comfort with Git, Linux, and structured experiment logging
• Practical experience training or fine-tuning deep models—vision-language, multimodal, or policy-learning setups—and interpreting loss curves, metrics, and checkpoints
• Familiarity with robot learning data formats (image-text-action pairs, trajectories, teleoperation logs) and basic kinematics / control interfaces
• Exposure to physics simulators for robotics (Isaac Sim / Isaac Lab, MuJoCo, Gazebo, or similar) and how synthetic data feeds learning pipelines
• Ability to scope work, execute methodically, and communicate results clearly in a fast-moving R&D environment
• High agency within defined priorities, good judgment on when to escalate blockers, and collaborative habits across ML and robotics disciplines

Nice to have

• Direct experience with VLA or embodied AI stacks (OpenVLA, RT-2, π₀, or team-equivalent baselines) and action representations for manipulation
• Background in cross-embodiment training, sim-to-real, domain randomization, or fleet-scale evaluation
• Exposure to reinforcement learning, imitation learning, or offline RL integrations adjacent to VLA policies
• Familiarity with ROS / ROS 2 deployment paths, CUDA, cluster or cloud job runners, and lightweight MLOps patterns
• Coursework or projects on dexterous manipulation, mobile manipulation, or contact-rich tasks
• Published or preprint research, strong Kaggle/robotics competition results, or notable open-source contributions in embodied AI

Who you are

• You want research that ships—models and evals that survive contact with real robots and operational constraints
• You are rigorous about reproducibility, versioning, and knowing what an experiment actually proved
• You are curious, practical, and comfortable learning from senior engineers and external advisors
• You care how VLA, simulation, and deployment compound into long-term product and fleet advantage
• You want to grow from scoped research engineering toward owning larger slices of our embodied AI roadmap

What we are looking to build

Research engineering foundations for production-oriented VLA stacks: functional model architectures and training pipelines that bridge high-level reasoning with reliable low-level control across embodiments, with validated data loops, sim benchmarks, and documented sim-to-real paths that turn experiments into assets the wider autonomy and platform teams can deploy and improve.