Helix AI Engineer, Pretraining

About the role

Figure is building autonomous humanoid robots designed to operate in real-world environments where unstructured human spaces are the norm. The Helix team creates the primary AI systems that drive these machines, and we are seeking an engineer to develop foundation models using multimodal data. In this role, you will own the end to end lifecycle of large scale model development from data strategy to deployed checkpoints. You will design the architecture and training regimes that allow robots to understand and act in the physical world. Your work will directly influence how future generations of embodied AI systems learn from interaction. This is a hands on position where your code and experiments will scale into production training jobs. You will be responsible for ensuring that our models generalize across diverse tasks and environments.

Key facts

What you'll do

• Architect and train foundation models using text, image, video, and robotics data to create a unified representation of the world.
• Create pretraining methods that enhance model reasoning, adaptability, and generalization across long horizon robotic tasks.
• Implement transformer architectures and other emerging model paradigms tailored for temporal and multimodal sequences.
• Research scaling laws, training dynamics, and optimal dataset mixtures to maximize learning efficiency and downstream performance.
• Manage distributed training pipelines across multi-node GPU clusters to ensure efficient and reliable experimentation.
• Coordinate with internal teams to incorporate pretrained models into the broader autonomy stack and production workflows.
• Build evaluation systems to assess cross-domain performance and reasoning in both simulated and real world scenarios.
• Assist with post training processes like alignment and fine tuning to refine behavior and safety characteristics.
• Design data ingestion frameworks that handle heterogeneous inputs from vision, language, and robot telemetry.
• Iterate on training objectives, loss functions, and optimization strategies to improve sample efficiency.
• Partner with robotics engineers to define task specific metrics that reflect real world success criteria.
• Conduct ablation studies and failure mode analysis to drive iterative improvements in model architecture.
• Maintain rigorous experiment tracking and documentation to ensure reproducibility and rapid knowledge transfer.
• Contribute to technical design documents and research discussions that shape the long term roadmap for Helix.

Requirements

• Professional background in training large-scale foundation models or LLMs with demonstrable experience on production workloads.
• Deep knowledge of transformer architectures and modern deep learning frameworks, including attention mechanisms and optimization techniques.
• Experience managing distributed training and optimization across clusters of GPUs with strong performance tuning skills.
• Proficiency in Python and PyTorch, with a portfolio of projects that highlight low level implementation and debugging.
• Ability to conduct rigorous experiments and iterate on model design based on empirical results and diagnostic signals.
• Software engineering skills focused on building scalable and reliable infrastructure that supports continuous training and deployment.
• Capacity to manage ambiguous technical challenges independently while communicating progress and risks clearly to stakeholders.
• Strong grasp of numerical computing, linear algebra, and probability to reason about model behavior and training stability.
• Familiarity with version control, containerization, and cloud compute resources essential for modern ML workflows.
• Willingness to follow detailed specifications and contribute to code reviews that maintain high quality standards.
• Commitment to safety and reliability principles when designing systems that interact with the physical world.
• Openness to mentorship and collaboration within a multidisciplinary team of researchers and engineers.
• Willingness to adhere to company policies, processes, and documentation practices without exception.
• Readiness to relocate to San Jose, CA and work onsite for the duration of the engagement as required.

Nice to have

• Prior work on frontier models at organizations like OpenAI, Anthropic, Google DeepMind, or xAI.
• Experience with vision language or vision language action models that bridge perception and control.
• Expertise in data curation and scaling law research to inform dataset growth and compute planning.
• Knowledge of alignment techniques such as RLHF or reward modeling to shape desired model behaviors.
• Familiarity with robotics or embodied AI deployment in real world scenarios and edge compute constraints.
• A documented record of research publications in AI, NLP, or multimodal learning that demonstrate technical depth.

Practical notes

Base salary range is $200,000 to $400,000. Total compensation and benefits are determined based on individual experience and skills, with details provided upon the extension of an offer.