Embodied AI Engineer

About RoboTensor

RoboTensor is a Physical AI research and development company building the infrastructure that turns robot foundation models into specialized physical capabilities.

We work across model adaptation, simulation, benchmarking, and evaluation to make general-purpose models capable of performing real tasks on real machines in real operating environments.

Our goal is simple: make robot intelligence measurable, improvable, and useful in the physical world.

About the Role

We are looking for an Embodied AI Engineer to own the end-to-end development and training of learned robot policies.

You will work across the full learning loop: generating and curating training data, building or adapting simulation environments, writing training pipelines, running experiments, evaluating policies, analyzing failures, and steadily improving model performance.

This is a highly hands-on role. You should be comfortable moving across data, simulation, models, training infrastructure, and evaluation rather than owning only one narrow piece of the stack.

Ultimately, your job is to make the policy better—iteration after iteration.

What You’ll Do

• Own policy training end to end, from training data generation through evaluation and deployment.
• Generate training data through demonstrations, simulation, synthetic data, existing datasets, and other sources.
• Build or adapt simulation environments when needed for training, data generation, or evaluation.
• Develop training and fine-tuning pipelines for vision-language-action models and other learned policies.
• Work with imitation learning, behavior cloning, reinforcement learning, offline learning, and related approaches.
• Design experiments that systematically improve policy performance.
• Build benchmarks and evaluation pipelines for measuring task success, robustness, generalization, and failure modes.
• Diagnose whether performance bottlenecks come from data, model architecture, training objectives, simulation, or evaluation.
• Turn policy failures into new data, experiments, and model improvements.
• Improve training efficiency, reproducibility, and experimental velocity.
• Implement promising ideas from current research and determine whether they improve real capabilities.
• Work with real systems to understand the gap between training performance and physical performance.

What We’re Looking For

• Strong machine learning and software engineering fundamentals.
• Experience training deep learning models with PyTorch, JAX, or similar frameworks.
• Experience building substantial parts of an ML training pipeline.
• Strong experimental intuition and the ability to systematically improve model performance.
• Ability to work comfortably across data, models, training infrastructure, and evaluation.
• Strong debugging skills across complex learning systems.
• Ability to independently take an ambiguous capability goal and turn it into a working training program.
• Comfortable operating in a small, fast-moving team with significant technical ownership.

Nice to Have

• Experience with embodied AI, robot learning, or autonomous systems.
• Experience with vision-language-action models or multimodal foundation models.
• Experience with imitation learning, reinforcement learning, or learning from demonstrations.
• Experience generating training data in simulation.
• Experience with physics simulators or digital-twin environments.
• Experience with distributed or large-scale model training.
• Experience deploying learned policies on physical systems.

What Success Looks Like

When a policy stops improving, you know how to investigate why.

You can determine whether the next improvement requires better data, a different environment, a training change, a model change, a better benchmark, or an entirely new experiment—and then build what is needed.

You can take a desired physical capability and work backward to determine what data to generate, how to train the model, how to measure success, and what to try next.