Founding Research Scientist

• Develop world action models and robot learning methods, including video-action learning, imitation learning, and reinforcement learning.
• Design controlled experiments to improve task success, generalization, robustness, and data efficiency.
• Train and post-train models on robot demonstrations and multimodal sensor data; analyze failures and identify the data needed to improve performance.
• Build reproducible training and evaluation workflows with clear baselines, metrics, ablations, and experiment records.
• Test policies on physical robots and collaborate on inference, control integration, and deployment.
• Help set research priorities and communicate results and technical tradeoffs clearly.

• At least two first-author publications at top-tier conferences in machine learning, computer vision, or robotics. This is a required qualification.
• Research experience in robot learning, machine learning, computer vision, or a closely related field.
• Strong Python and PyTorch skills, including experience implementing, training, and debugging learning systems.
• Ability to design rigorous experiments, interpret results, and translate research ideas into working code.
• Experience with robotic systems or a strong interest in hands-on experimentation with physical hardware.
• A Ph.D. in a relevant field or equivalent research experience.

• World action models, video generation, diffusion policies, or vision-language-action models.
• Manipulation, teleoperation, reinforcement learning, or sim-to-real transfer.
• Large-scale training, multimodal datasets, and systematic policy evaluation.

• Location: South San Francisco, California; on-site.
• Employment: Full-time.
• Annual base salary: $200,000–$350,000 USD.

Help build models that learn real jobs and robots that perform them reliably. Include links to at least two qualifying first-author conference publications and describe your contributions. Code, project write-ups, and demonstrations are welcome too.

Verne is building physical AI to make physical work as scalable as software and bring general-purpose robots to every business. We develop robots, foundation models, and the tools that turn them into reliable industrial deployments.

Our open-weight world action model, Verne V0.5 https://vernerobotics.com/research/verne-v0-5, brings together video and action learning. Alongside it, we build data and training systems, a fine-tuning platform, and inference and control software. Our work is grounded in real deployments: learning new tasks efficiently, adapting to different environments, and operating reliably in production.