Robotics Foundation Model Engineer

Lightwheel; California (MO); On-site; USD 140,000 - 200,000; Full time

Job summary

Lightwheel is seeking a research-focused engineer to advance vision-language-action models for real-world robotics. The role involves designing end-to-end training loops, coordinating with simulation and deployment teams, and driving robust policy deployment across robot platforms.

Ideal candidates hold a Master's or PhD with strong Python/PyTorch skills and experience in RL, multimodal models, and robot learning.

Qualifications

• Master's or PhD in machine learning, robotics, computer vision, control, or a related field preferred.
• Deep experience in at least one of the following areas: VLA models, robot learning, reinforcement learning, imitation learning, or multimodal foundation models.
• Proficiency in Python and PyTorch; experience with CUDA/JAX, distributed training, or large-scale data pipelines is a plus.
• Experience with real robots, simulation-based training, teleoperation/human data, or model deployment.
• Able to clearly articulate personal contributions, training data, baselines, ablations, failure cases, and the limits of reported results.

Responsibilities

• Develop vision-language-action (VLA) models and policies, Diffusion Policies, behavior cloning, offline/online reinforcement learning, and related training methods.
• Design unified representations, data mixture strategies, and quality evaluation systems across human, robot, and simulation data.
• Build an end-to-end training loop spanning pretraining, supervised fine-tuning (SFT), preference optimization, policy evaluation, and real-robot rollouts.
• Research embodiment adaptation across robot platforms, grippers, and sensor configurations.
• Collaborate with simulation, world model, data, and deployment teams to move models from experimentation into real-world robotic systems.
• Establish quantitative metrics for model generalization, long-tail failures, and deployment stability.

Skills

Python

PyTorch

Reinforcement learning

Multimodal models

Robot learning

Education

Master's or PhD in ML/Robotics/CS

Tools

CUDA

JAX

C++

distributed training

large-scale data pipelines

Job description

Responsibilities

• Develop vision-language-action (VLA) models and policies, Diffusion Policies, behavior cloning, offline/online reinforcement learning, and related training methods.
• Design unified representations, data mixture strategies, and quality evaluation systems across human, robot, and simulation data.
• Build an end-to-end training loop spanning pretraining, supervised fine-tuning (SFT), preference optimization, policy evaluation, and real-robot rollouts.
• Research embodiment adaptation across robot platforms, grippers, and sensor configurations.
• Collaborate with simulation, world model, data, and deployment teams to move models from experimentation into real-world robotic systems.
• Establish quantitative metrics for model generalization, long-tail failures, and deployment stability.

Qualifications

• Master's or PhD in machine learning, robotics, computer vision, control, or a related field preferred.
• Deep experience in at least one of the following areas: VLA models, robot learning, reinforcement learning, imitation learning, or multimodal foundation models.
• Proficiency in Python and PyTorch; experience with C++, distributed training, CUDA/JAX, or large-scale data pipelines is a plus.
• Experience with real robots, simulation-based training, teleoperation/human data, or model deployment.
• Able to clearly articulate personal contributions, training data, baselines, ablations, failure cases, and the limits of reported results.

Preferred Qualifications

Experience with ALOHA, RoboTwin, Isaac Sim/Isaac Lab, MuJoCo, dexterous hands, tactile sensing, multimodal action representations, long-horizon tasks, or automated policy improvement.