VLA/WAM Research Engineer

Flexion Robotics•San Francisco, CA

About The Position

Flexion is building the intelligence layer for humanoid robots, with the goal of accelerating the transition from prototypes to real-world deployment. The VLA/WAM Research Engineer will develop foundation models for humanoid robots across model architecture, large-scale pretraining, adaptation, and post-training. The role spans the model lifecycle: designing vision-language-action (VLA) and World-Action Model (WAM) architectures; training foundation models on diverse embodied datasets; developing efficient methods to adapt models to new tasks and embodiments; and using robot experience to improve model capabilities. The position suits researchers and engineers with depth in one or more areas such as multimodal model architecture, large-scale pretraining, world models, robot learning, or post-training and continual learning, who are eager to work across the broader challenge of building general-purpose robot foundation models.

Requirements

Strong Python and PyTorch skills, with hands-on experience developing and training modern deep-learning models.
Research or engineering experience in one or more of: VLA/VLM architectures, multimodal foundation models, world models, large-scale pretraining, robot learning, or reinforcement learning/post-training.
Experience owning substantial model-development or training experiments, including architecture choices, training objectives, optimization, and evaluation.
Solid understanding of modern deep-learning architectures, representation learning, and how data and training objectives relate to model performance.
Ability to design rigorous experiments, diagnose training failures, and use empirical results to guide research decisions.
Familiarity with large-scale datasets, model-training workflows, and generalization challenges across tasks, environments, or embodiments.
Ability to critically evaluate recent research and translate promising ideas into working models and experiments.
Comfort working across model architecture, training, and adaptation while bringing particular expertise in at least one of these areas.

Nice To Haves

Experience training VLA, VLM, diffusion, flow-matching, video/world models, or other multimodal generative architectures.
Experience with distributed training, multi-node GPU systems, large-scale data mixtures, and foundation-model training optimization.
Experience with reinforcement learning, offline RL, reward/value learning, specialist distillation, or continual learning.
Experience with predictive world models, video generation, latent dynamics, or self-supervised representation learning.
Experience with multi-embodiment robot learning, action representations, or adapting policies across different robot platforms.
Experience working with teleoperation, egocentric video, simulation, autonomous rollouts, or other large-scale embodied datasets.
Experience evaluating learned policies on real robots, particularly humanoids or dexterous manipulation systems.

Responsibilities

Design and develop multimodal foundation models that combine vision, language, robot state, actions, and temporal context for general-purpose robotic control.
Develop and execute large-scale pretraining across diverse robot embodiments, teleoperation, egocentric human data, and simulated experience.
Improve data mixtures, training objectives, optimization strategies, and distributed training workflows to deliver increasingly capable foundation models regularly.
Develop predictive world and action models for representation learning, future-state prediction, planning, and robot policy improvement.
Develop efficient methods to adapt pretrained models to new tasks, environments, embodiments, and capabilities.
Improve models using real-world and simulated robot experience, including imitation learning, reinforcement learning, reward/value-guided learning, and specialist-to-generalist distillation.
Develop continual-learning methods using autonomous rollouts, failures, interventions, and recoveries while retaining existing capabilities and avoiding catastrophic forgetting.
Design rigorous experiments, evaluate models in simulation and on real robots, and use results to guide architecture and training decisions.