Research Engineer
At a glance
Mid-level AI Research role at Hedra. San Francisco · full-time · $175,000–$275,000 base.
aimlroles summary, based on the employer's posting.
What you'll do
- Build and run training pipelines for action-conditioned world models and vision-language-action models
- Improve model training through fine-tuning, reinforcement learning and large-scale multimodal methods
- Create simulation-based training and evaluation data, including environment setup and sim-to-real approaches
- Develop distributed training systems with PyTorch, FSDP and DeepSpeed, and handle video, sensor and action data
- Assess models with benchmark datasets and deployment results, while adapting them with industrial partners
What you bring
- You have trained large generative models before, such as video, multimodal or action-conditioned systems
- You can use PyTorch and distributed training tools such as FSDP or DeepSpeed hands-on
- You bring strong machine learning, optimization and large-scale data processing fundamentals
- You have a BS, MS or PhD in computer science, machine learning, robotics or a related field
Who this fits
This role suits an engineer who wants to build and evaluate generative models for physical systems, with work spanning simulation and industrial applications. The position is based in San Francisco.
From the employer
Overview:
Hedra is a pioneering generative modeling company — first models to market — now building a Physical AI team to bring these models to real-world industry and economy use cases. As a Research Engineer on our Physical AI team, you will lead pre-training and post-training on action-conditioned world models, working hand-in-hand with industrial partners to close the loop between generative AI and physical systems. This is not a black-box applied role: your work will be published, your infrastructure will be serious, and your impact will be direct. If you want to work at the frontier of generative modeling and physical AI, this is the team.
Responsibilities:
Design, implement, and run pre-training and post-training pipelines for action-conditioned world models and vision-language-action (VLA) models
Develop and refine training methodologies, including fine-tuning, reinforcement learning, and large-scale multimodal learning
Design and generate training and evaluation datasets from simulation, including environment setup, domain randomization, and sim-to-real transfer strategies
Build distributed training infrastructure using PyTorch, FSDP, and DeepSpeed
Work with multimodal data pipelines involving video, sensory inputs, and action sequences
Evaluate model performance using both benchmark datasets and real-world deployment metrics
Contributions research publications a plus
Collaborate with industrial partners to adapt generative models for real-world physical AI applications
Qualifications:
Experience with pre-training or post-training on large generative models (video, multimodal, or action-conditioned)
Hands-on proficiency with PyTorch and distributed training frameworks (FSDP, DeepSpeed)
Strong fundamentals in machine learning, optimization, and large-scale data processing
Familiarity with VLMs, VLAs, or world models
Background in robotics, embodied AI, or sim-to-real transfer is a plus
Experience with video understanding or temporal reasoning is a plus
BS/MS/PhD in Computer Science, Machine Learning, Robotics, or a related field
Benefits:
Competitive compensation and equity
401k (no match)
Healthcare (Silver PPO Medical, Vision, Dental)
Lunch and snacks at the office
We encourage you to apply even if you don't fully meet all the listed requirements; we value potential and diverse perspectives, and your unique skills could be a great asset to our team.