Senior AI Engineer ( Annotation & VLA Models

We build robot learning systems for humanoid robots. We collect teleoperation and egocentric data, label it, and use it to train vision-language-action (VLA) models. We train and test these models in simulation with reinforcement learning (RL), in simulation with human corrections in VR, and on real robots. We have our own annotation model for robot data, and a data catalogue that keeps episodes in the LeRobot format.

The core team is small and senior, and it does the technical work together. Our goal is to find gaps in the current state-of-the-art robotics methods. Then we propose improvements to these methods.

Lead the VLA and annotation work of the team, with help from the other engineers. Most of your time goes to the VLA models: with the team, select them, fine-tune them, make them better and test them. You also own the annotation of our robot data. Our annotation model is already in production, so this part of the role is about label quality and about labels that make the policies better, not about building annotation tools from zero.

Use published VLA research to make policies that do useful tasks on our robots. Repeat published experiments with the team and compare the results. Design and test new methods for post-training, data efficiency and inference speed. When the team grows, lead a team in this area.

• With the team, select the base VLA models for our robots and tasks. Compare different models on the same tasks and data. Examples of these models are pi0.5, GR00T N1.7 and Psi
  0.
• Fine-tune VLA models on teleoperation demonstrations from simulation and from real robots. Use the code base of each model. Examples of code bases are LeRobot, openpi and Isaac-GR00T.
• Do RL post-training of VLA models in simulation, for example with RECAP-style advantage conditioning in BiGym on Mu Jo Co . Measure each new policy against its baseline.
• Use human corrections to make the policies better. For example, use VR corrections in simulation and HG-DAgger with the RaC rules on real robots.
• Define the evaluation protocol for simulation and for real robots. Use fixed test sets, blind A/B trials and statistical tests, for example Wilson intervals and McNemar tests. Find the causes of failures in rollouts, logs and videos.
• Run the policies on real robots with the robotics team. For example, use the Unitree G1 with Unitree SDK2 and ROS 2.
• Make the inference of the policies fast enough for real-time control. For example, use TensorRT, mixed precision and fewer integration steps in the action expert.
• Measure how many demonstrations are necessary for each task. Test methods that get the same result with fewer demonstrations. For example, use generated data from Humanoid Mimic Gen .
• Test new research ideas on our robots. Examples of these ideas are cross-embodiment transfer and egocentric human video in pre-training. Help the simulation engineer test world models as policy evaluators.
• Write a technical report on each experiment and on each new method.

• Own the annotation model and the annotation QA for our robot data. Work with the annotation team, which does the daily operations.
• Define the label schema for VLA training: language instructions, subtask segmentation, success and failure labels, and hand-pose and action QA.
• Do data ablations to find which data and which labels improve the policies most.
• Use the results of VLA experiments to decide what to label next and what data to capture next. Send these requests to the annotation team and to the data capture team.
• Improve the annotation model where label errors cause policy failures. For example, use vision-language models (VLMs) for automatic labels and human review for the difficult cases.
• Keep the labels versioned in the data catalogue, so that each training run uses a known dataset.

• 5 or more years of experience with deep learning, and 3 or more years with robot policies or embodied AI.
• Experience with the fine-tuning of a VLA model or an imitation policy, and with the operation of that policy on a real robot. Examples of these policies are pi0, pi0.5, GR00T, OpenVLA, ACT and Diffusion Policy.
• Strong knowledge of robot imitation learning, for example behavior cloning and DAgger-style corrections. Good knowledge of reinforcement learning for robots.
• Expert knowledge of PyTorch. Experience with the fine-tuning of large transformer models on multi-GPU systems, with mixed precision and LoRA.
• Good…