Research Scientist (Vision-Language-Action Models)

About the role

The AI Robot Association (AIRoA) is launching a project to collect one million hours of humanoid robot operation data and use it to train powerful Vision-Language-Action (VLA) models, with datasets, models and benchmarks made open to researchers. It is recruiting a Research Scientist to work on multimodal learning, tactile-rich manipulation, sim-to-real transfer and large-scale benchmarking, with access to hundreds of humanoid robots, GPU clusters and simulators. The project is a national initiative supported by 20.5 billion yen from NEDO.

Key responsibilities

• Design and implement data preprocessing pipelines for multimodal robot datasets
• Train VLA models using supervised learning, RL, fine-tuning, RLHF and training from scratch
• Develop and evaluate models in simulation and on physical robots
• Improve training robustness and efficiency through algorithmic innovation
• Analyze model performance and propose enhancements based on empirical results
• Deploy VLA models onto real humanoid and mobile robotic platforms
• Publish research in top-tier conferences such as NeurIPS, CoRL, CVPR

Required profile

• MS degree with 3+ years of industry experience, or PhD in Computer Science, Electrical Engineering or a related field
• At least one first-author publication in a top-tier conference (CoRL, ICML, CVPR, NeurIPS, IROS, ICLR, ICCV, ECCV)
• Experience with open-ended learning, reinforcement learning and frontier methods for training LLMs/VLMs/VLAs such as RLHF and reward function design
• Experience with simulators or real-world robots
• Knowledge of recent advances in large-scale machine learning research

Required skills

• Deep learning frameworks such as PyTorch
• Preferred: robot learning, impedance, adaptive, force control or MPC, tactile sensing, ROS, and simulators such as MuJoCo, PyBullet, Isaac Sim