Your Mission

As Research Scientist / Engineer — Large-scale Training, you will develop the architectures and training methods that make UMA’s robotic foundation models more capable and more efficient to train. You will investigate how model design, optimization, and hardware-aware implementation work together to deliver better humanoid robot performance within a given compute budget.

Your primary focus will be pretraining, with work extending into mid-training and post-training. You will explore efficient multimodal architectures, faster training recipes, and compute-efficient scaling strategies, using controlled experiments to determine which improvements translate into stronger capabilities on real robots.

We already have a working training stack and a strong team behind it. Working closely with AI researchers and Training Infrastructure engineers, you will take ideas from architectural changes and kernel-level prototypes through training experiments to larger-scale validation. Your findings will shape what we train, how we train it, and how we allocate compute.

Key responsibilities:

• Research efficient model architectures for robotic foundation models, including attention mechanisms, multimodal fusion, visual and action tokenization, sequence compression, and dense versus sparse or mixture-of-experts designs.

• Optimize training recipes and convergence through choices of optimizers, learning-rate schedules, initialization, normalization, batch sizes, and numerical precision. Measure improvements in time and compute required to reach a target capability.

• Co-design architectures and compute kernels: identify expensive operations, explore hardware-friendly alternatives, and prototype fused or custom kernels using tools such as Triton or CUDA where useful.

• Investigate efficient attention and sequence processing, reducing the cost of learning from video, temporal context, and action sequences while preserving useful information.

• Develop scaling laws and compute-allocation studies to understand how performance changes with model size, training data, and compute. Use smaller experiments to guide larger runs, and validate where those predictions hold.

• Research training acceleration techniques, such as progressive increases in resolution or context length, selective freezing, parameter sharing, and lower-precision computation, evaluating their effects on stability and final performance.

• Run controlled ablations and performance benchmarks that distinguish faster execution from better learning efficiency, tracking wall-clock time, compute, memory, convergence, and downstream capability.

• Diagnose training dynamics and numerical instabilities, including gradient behavior, sensitivity to precision, loss spikes, and changes that emerge as models or batch sizes scale.

• Evaluate how pretraining improvements transfer through mid-training and post-training to real-robot success, generalization, robustness, and adaptation efficiency.

• Work with Training Infrastructure engineers to integrate validated methods into scalable, reproducible training systems, and with Data researchers to study interactions between model capacity, data mixtures, and training duration.

What You Bring to the Table

• Previous experience in training large-scale frontier foundation models such as LLMs, Video Generation Or Vision Encoders

• Strong experience in deep learning research, efficient model training, foundation models, or ML systems, demonstrated through research projects, publications, deployed systems, or open-source work.

• Strong Python and PyTorch, with hands-on experience implementing architectures and running substantial training experiments.

• A deep understanding of model architectures, optimization, and training dynamics, and the ability to explain why a change affects convergence, stability, or computational cost.

• Experience improving training efficiency, whether through architecture design, optimization methods, numerical precision, or hardware-aware implementation.

• The ability to reason about compute, memory bandwidth, tensor shapes, and GPU execution, and connect profiling results to research decisions.

• Strong experimental judgment: meaningful baselines, fair compute comparisons, reproducible results, and careful interpretation of scaling trends.

• Interest in connecting efficient training research to multimodal learning and real-world robot capabilities.

• Thrive in a hands-on, fast-paced startup: autonomous, rigorous, collaborative, and comfortable moving between research and implementation.

• Bonus: Triton, CUDA, fused kernels, efficient attention, or compiler-based optimization.

• Bonus: scaling laws, compute-optimal training, mixture-of-experts models, or large-scale multimodal pretraining.

• Bonus: distributed training, low-precision training, vision-language-action models, or robot learning.

• Bonus: open-source contributions, reproducible research implementations, or technical writing.

We value deep expertise in one or more of these areas, alongside the curiosity to work across them. UMA is an inclusive workplace that values exceptional researchers and builders over perfect pedigrees. Whatever your background, identity, or journey, if you don’t meet every criterion but believe you can have an outsized impact here, we strongly encourage you to apply.

Make every unit of compute count toward more capable robots.