Staff Machine Learning Engineer, VLA Training

Nidus is building the intelligence layer that makes that possible. The team comes from the intersection of defense technology and AI, with deep experience across ML, robotics, product development, and defense, and a track record of shipping products to DoD customers.

THE OPPORTUNITY

Nidus is building intelligent robots that can perceive the world, understand instructions, and perform useful physical tasks in dynamic, real-world environments. As Staff Machine Learning Engineer, you will own major parts of the model development lifecycle — from dataset construction and training recipes to distributed training infrastructure, evaluation, and on-robot validation.

This is a hands-on role at the intersection of machine learning research, large-scale systems, and robotics. You will work closely with researchers, robotics engineers, data teams, and operators to turn new modeling ideas into reliable robot capabilities. The right person is equally comfortable investigating why a policy failed on a robot, designing the next training experiment, and improving the infrastructure required to run that experiment efficiently at scale.

WHAT YOU’LL DO

• Train and improve vision-language-action models for robotic manipulation and other embodied tasks


• Develop model architectures and training recipes spanning imitation learning, behavior cloning, transformer- and diffusion-based policies, multimodal pretraining, and post-training


• Build scalable pipelines for pretraining, fine-tuning, evaluation, checkpointing, and model release


• Train models across multi-node GPU clusters while improving utilization, throughput, stability, and cost efficiency


• Design data mixtures, sampling strategies, augmentations, and curriculum approaches for large, heterogeneous robot datasets


• Develop data loaders and preprocessing systems for synchronized video, language, robot state, actions, and other sensor modalities


• Create reproducible experimentation systems including configuration management, dataset and model versioning, experiment tracking, and automated regression testing


• Define offline and on-robot metrics that measure task success, generalization, robustness, latency, safety, and failure modes


• Build tools for inspecting trajectories, visualizing model behavior, comparing experiments, and diagnosing data or training failures


• Run structured experiments on physical robots and use the results to guide model, data, and infrastructure improvements


• Partner with data collection teams to identify coverage gaps, improve demonstration quality, and prioritize collection based on model performance


• Make technical decisions across model architecture, data, compute, and evaluation, and communicate the associated tradeoffs clearly


• Mentor other engineers and establish strong engineering practices for the ML codebase

YOU SHOULD HAVE

• 5+ years of professional experience in machine learning, robotics, computer vision, or a closely related field


• Strong experience training modern deep learning models using PyTorch, JAX, or a comparable framework


• Experience with transformer-based models, multimodal models, generative models, or learned control policies


• Experience building and operating distributed training pipelines on multi-GPU or multi-node accelerator clusters


• Experience with large-scale datasets, high-throughput data loading, experiment tracking, and reproducible ML workflows


• Strong understanding of optimization, model architecture, data quality, evaluation methodology, and experimental design


• Ability to debug failures across the full stack, from input data and training dynamics to inference and robot behavior


• Experience independently owning technically ambiguous projects and delivering working systems

NICE TO HAVE

• Experience developing vision-language-action models, vision-language models, or robotics foundation models


• Experience with imitation learning, reinforcement learning, behavior cloning, diffusion policies, action tokenization, or action-conditioned world models


• Experience training policies using data from multiple robot embodiments, task domains, or sensor configurations


• Familiarity with large-scale pretraining, post-training, parameter-efficient fine-tuning, distillation, or model adaptation


• Experience optimizing distributed training using techniques such as FSDP, tensor or pipeline parallelism, mixed precision, gradient checkpointing, or sharded data loading


• Familiarity with robotics middleware and simulation environments such as ROS/ROS 2, MuJoCo, Isaac Sim, or Gazebo


• Experience deploying and evaluating learned policies on physical robotic systems


• Publications or meaningful open-source contributions in machine learning, robotics, or computer vision