Senior AI Engineer (Annotation & Vision-Language-Action Models)

About the Team

We develop robotic learning systems tailored for humanoid robots. Our process includes gathering teleoperation and first-person perspective data, annotating it, and leveraging it to train vision-language-action (VLA) models. These models are trained and evaluated through simulations employing reinforcement learning (RL), simulations with human corrections via virtual reality (VR), and testing on physical robots. We utilize a proprietary annotation model for robot data and maintain a data catalogue formatted in LeRobot standard. Our core team is small, senior, and collaboratively performs technical tasks. Our mission focuses on identifying shortcomings in current state-of-the-art robotics approaches and enhancing these techniques.

Role Overview

Lead the team’s efforts regarding VLA models and annotation activities, assisted by other engineers. Your primary focus will be on VLA models: selecting, fine-tuning, improving, and validating them alongside the team. You will also oversee the annotation quality of robot data, focusing on label accuracy and effectiveness for policy improvement rather than developing annotation tools from scratch.

Key Responsibilities

• Collaborate to choose base VLA models suitable for robots and tasks, benchmarking models such as pi0.5, GR00T N1.7, and Psi0 on consistent datasets.
• Fine-tune VLA models using teleoperation demonstrations from both simulations and physical robots, utilizing relevant codebases like LeRobot, openpi, and Isaac-GR00T.
• Apply reinforcement learning post-training techniques (e.g., RECAP-style advantage conditioning in BiGym on MuJoCo) and quantify policy improvements over baselines.
• Incorporate human corrections, including VR simulation adjustments and HG-DAgger with RaC rules on physical hardware, to enhance policy performance.
• Develop and implement evaluation protocols for both simulation and real robot tests, employing fixed test sets, blind A/B testing, and statistical analyses like Wilson intervals and McNemar tests. Investigate failure reasons using rollouts, logs, and videos.
• Execute policies on real robots in collaboration with the robotics team, including devices like Unitree G1 using Unitree SDK2 and ROS 2.
• Optimize policy inference speed for real-time control using techniques such as TensorRT, mixed precision computing, and reducing integration steps of action experts.
• Assess the minimum number of demonstrations required per task and explore methods to maintain performance with fewer demonstrations, utilizing tools such as HumanoidMimicGen for data generation.
• Test novel research concepts on robots, like cross-embodiment transfer and egocentric human video pre-training, and support simulation engineers with policy evaluation using world models.
• Produce detailed technical documentation for each experiment and developed method.
• Manage the annotation model and quality assurance for robot datasets in coordination with the annotation team handling daily operations.
• Define labeling schemas for VLA training encompassing language instructions, subtask segmentation, success/failure tags, and hand-pose/action quality checks.
• Conduct data ablations to determine which datasets and labels enhance policy quality the most.
• Direct labeling priorities and data collection efforts based on VLA experiment results, collaborating with annotation and data capture teams.
• Refine the annotation model to rectify label errors impacting policies, incorporating vision-language models for automatic labeling plus human reviews for complex cases.
• Maintain versioned labels in the data catalogue to ensure reproducibility for each training iteration.

Required Skills and Experience

• A minimum of five years in deep learning with at least three years dedicated to robot policies or embodied AI.
• Hands-on expertise in fine-tuning VLA models or imitation policies and deploying these on physical robots. Familiarity with policies such as pi0, pi0.5, GR00T, OpenVLA, ACT, and Diffusion Policy.
• In-depth understanding of robot imitation learning techniques, including behavior cloning and DAgger-style human corrections, as well as reinforcement learning applied to robotics.
• Advanced proficiency with PyTorch, especially in fine-tuning large transformer models using multi-GPU infrastructure, mixed precision, and parameter-efficient approaches like LoRA.
• Experience working with at least one open-source VLA model codebase such as LeRobot, openpi, or Isaac-GR00T.
• Strong background managing large multimodal and video datasets: handling data pipelines, quality assurance, and dataset version control.
• Capability in designing robust policy evaluation methodologies, including fixed test datasets, sufficient trial numbers, and statistical confidence metrics.
• Practical knowledge of accelerating inference to achieve real-time control using tools like TensorRT, ONNX, quantization, and mixed precision adjustments.
• Track record of senior-level contributions including robotics or machine learning publications, widely adopted open-source projects, or deployed production-grade robotic systems.
• Proven ability to communicate technical findings and potential risks clearly to engineering teams and management.

Desirable Experience

• Background in annotating robot or video data, including designing label schemas, QA processes, and measuring annotator agreement.
• Exposure to model-assisted labeling techniques, such as episode segmentation and linguistic annotation using vision-language models.
• Familiarity with humanoid robots and dexterous hand hardware/software like Unitree G1, Dex3-1 hand, Unitree SDK2, and ROS 2.
• Experience with offline reinforcement learning or advantage-weighted post-training methods like RECAP and AWR for VLA models.
• Expertise evaluating robot policies in simulation platforms such as MuJoCo, BiGym, Isaac Sim, or Isaac Lab.
• Knowledge of VR-based teleoperation (e.g., xr_teleoperate or WiVRn) and human-in-the-loop correction methodologies such as RaC-style recovery data.
• Understanding of whole-body humanoid control systems, for example NVIDIA GR00T-WholeBodyControl.
• Involvement with teleoperation datasets, egocentric human video analysis, cross-embodiment knowledge transfer, or robotic world model applications.
• Skills in JAX for openpi and inference deployment on NVIDIA Jetson platforms.
• Experience leading small technical teams.
• Possession of a PhD in robotics or machine learning, or equivalent research qualifications.