Research Engineer, Robot Learning

About Robotensor

Robotensor is building general-purpose intelligence for robots.

We work on robot foundation models, multimodal learning, robot data, simulation, and real-world deployment. Our goal is to build models that learn from large and diverse datasets, generalize across tasks and environments, and translate learned representations into reliable physical behavior.

About the Role

We are looking for a Research Engineer, Robot Learning to train and improve large-scale robot learning models.

You will work across the model training lifecycle: data preparation, pre-training, post-training, fine-tuning, experimentation, evaluation, and deployment on physical robots.

This is a hands-on research engineering role. You will run training at scale, investigate why models succeed or fail, implement new learning methods, and continuously improve model capability through changes to data, objectives, architectures, and training recipes.

What You’ll Do

• Train robot learning and Vision-Language-Action models on large-scale multimodal datasets.
• Own pre-training, post-training, fine-tuning, and adaptation pipelines.
• Develop training recipes for robot foundation models.
• Build and optimize datasets combining robot trajectories, demonstrations, video, language, proprioception, and action data.
• Experiment with data mixtures, sampling strategies, augmentation, filtering, and curriculum design.
• Develop and test training objectives for imitation learning, behavior cloning, reinforcement learning, and offline learning.
• Implement and experiment with modern robot policy architectures.
• Run large-scale experiments and systematic ablations.
• Analyze training dynamics, model behavior, regressions, and failure modes.
• Improve model generalization across tasks, objects, environments, and robot embodiments.
• Develop methods for efficient fine-tuning and adaptation to new tasks and robots.
• Optimize distributed training for throughput, stability, and GPU utilization.
• Build reproducible training pipelines and experiment infrastructure.
• Work closely with data, robotics, simulation, and evaluation engineers.
• Take models from training runs through evaluation and deployment on physical robots.

Problems You May Work On

Pre-Training

How should large and heterogeneous robot datasets be combined to produce general reusable representations and behaviors?

Post-Training

How can a general model be efficiently adapted to a specific embodiment, environment, or task?

Data Mixtures

What combination of robot trajectories, human demonstrations, video, language, simulation, and other data produces the strongest downstream capabilities?

Cross-Embodiment Learning

How can training on one robot improve performance on another robot with different sensors, kinematics, or action spaces?

Sample Efficiency

How can robots acquire reliable new behaviors from fewer demonstrations and fewer physical interactions?

Scaling

How does robot model performance change with model size, dataset size, diversity, compute, and training duration?

Robustness

How do we train policies that continue to work when the real environment differs from the training distribution?

What We’re Looking For

• Strong Python and PyTorch experience.
• Experience training deep learning models end-to-end.
• Strong understanding of modern deep learning architectures and optimization.
• Experience developing and debugging large training pipelines.
• Strong experimental methodology and ability to run meaningful ablations.
• Experience working with large multimodal datasets.
• Ability to diagnose whether a model limitation comes from data, optimization, architecture, or training objectives.
• Ability to implement research ideas and evaluate them quickly.

Relevant experience may include:

• Robot learning
• Vision-Language-Action models
• Imitation learning
• Reinforcement learning
• Offline RL
• Multimodal foundation models
• Transformers
• Diffusion or flow-based policies
• World models
• Representation learning
• Distributed training

Nice to Have

• Experience training robot foundation models.
• Experience with large robot trajectory datasets.
• Experience training policies for manipulation.
• Experience with multi-robot or cross-embodiment datasets.
• Experience with pre-training and post-training large models.
• Experience with distributed GPU training.
• Experience with JAX or large-scale PyTorch training systems.
• Experience deploying learned policies on physical robots.
• Publications or strong open-source work in robot learning, embodied AI, machine learning, or related areas.

What Success Looks Like

Training runs are reproducible and scalable.

New datasets and training recipes lead to measurable capability improvements.

Models generalize better to new tasks, objects, environments, and embodiments.

New robot capabilities require progressively less task-specific data.

Training failures can be diagnosed quickly rather than treated as black-box outcomes.

And improvements measured during training translate into better behavior on real robots.