Senior AI Engineer ( Annotation & VLA Models)

Team and role

VLA models and policy evaluation

• Select and compare base VLA models for the team's robots and tasks; examples include pi0.5, GR00T N1.7, and Psi0.
• Fine-tune models on simulation and real-robot teleoperation demonstrations using codebases such as LeRobot, openpi, and Isaac-GR00T.
• Perform simulation RL post-training, compare policies with baselines, and use human corrections in simulation and on real robots.
• Define evaluation protocols using fixed test sets, blind A/B trials, and statistical tests; investigate failures using rollouts, logs, and videos.
• Run policies on real robots, including Unitree G1 systems using Unitree SDK2 and ROS 2.
• Optimize inference for real-time control using techniques such as TensorRT, mixed precision, and fewer action-expert integration steps.
• Measure demonstration needs, test data-efficiency methods, explore research ideas on robots, and help evaluate world models.
• Write technical reports on experiments and new methods.

Annotation and training data

• Own the annotation model and robot-data annotation QA with the annotation team.
• Define VLA training labels for language instructions, subtask segmentation, success and failure, and hand-pose and action QA.
• Run data ablations to identify which data and labels improve policies.
• Use experiment results to guide future labeling and data capture requests.
• Improve annotation where label errors cause policy failures, including VLM-generated labels and human review of difficult cases.
• Maintain versioned labels in the data catalogue so training runs use known datasets.

Necessary skills and experience

• At least 5 years of deep-learning experience and 3 or more years with robot policies or embodied AI.
• Experience fine-tuning a VLA model or imitation policy and operating it on a real robot.
• Strong robot imitation-learning knowledge, including behavior cloning and DAgger-style corrections; good knowledge of robot reinforcement learning.
• Expert PyTorch knowledge; experience with large-transformer fine-tuning on multi-GPU systems, mixed precision, and LoRA.
• Good knowledge of one open-source VLA codebase, such as LeRobot, openpi, or Isaac-GR00T.
• Experience with large multimodal or video training datasets, including pipelines, data-quality checks, and dataset versioning.
• Experience designing robot-policy evaluations with fixed test sets, sufficient trials, and confidence intervals.
• Experience with fast inference for real-time control; examples include TensorRT, ONNX, quantization, and mixed precision.
• Evidence of senior work through robotics or machine-learning papers, widely used open-source code, or production robot systems.
• Ability to communicate results and risks clearly to engineers and managers.

Helpful experience

• Robot or video annotation, label-schema design, annotation QA, and annotator agreement are helpful.
• Model-assisted labels, humanoid robots and dexterous hands, offline RL, robot simulation, VR teleoperation, whole-body control, teleoperation datasets, egocentric video, cross-embodiment transfer, or robot world models are helpful.
• JAX for openpi, NVIDIA Jetson inference, and leading a small technical team are helpful.
• A PhD in robotics or machine learning, or equivalent-level research experience, is helpful.