The role:
Own the loop from raw robot experience to a model running on real hardware.
Munari observes what robots see, sense, decide, and do. We want to use that data to detect abnormal behavior, understand failures, evaluate policies, identify recurring failure modes, and help robotics teams improve the next version of their system.
This role sits deliberately between research engineering and production machine learning. You might train a multimodal model, redesign the dataset, optimize its inference path, deploy it to an edge device, and then test it against a real robot in the same week.
We are not looking for research that ends at a chart. We are looking for research that becomes a working system.
What you will work on:
Train multimodal models over video, language, proprioception, motor currents, force, tactile, IMU, telemetry, trajectories, and policy traces.
Build models for anomaly detection, episode scoring, failure classification, incident understanding, representation learning, and behavioral evaluation.
Integrate, evaluate, adapt, or fine-tune vision-language-action models and other robot policies across different embodiments.
Develop training and evaluation methods for limited, noisy, heterogeneous, and highly imbalanced real-world robot data.
Turn failures, operator takeovers, recoveries, and successful episodes into evals, labeled datasets, and training signals.
Distill, quantize, prune, compile, and optimize models for deployment on constrained edge hardware.
Design benchmarks, ablations, fault-injection experiments, and metrics that tell us whether a model will actually improve real-world operation.
Work closely with the infrastructure team on distributed training and data systems, and with the edge team on low-latency inference.
Validate ideas in simulation and then close the loop on physical robots.
You may be a strong fit if:
You have trained modern vision, transformer, diffusion, generative, or multimodal models.
You are highly productive in PyTorch, JAX, or a comparable ML stack.
You understand both model behavior and the systems required to train and deploy models reliably.
You have experience with distributed training, large datasets, inference optimization, or GPU performance.
You can move comfortably between research code, data pipelines, production services, and edge deployment.
You are excited by messy real-world data and do not require a perfectly clean benchmark before beginning.
You have strong experimental judgment and can distinguish genuine improvement from an attractive demo.
Experience with imitation learning, reinforcement learning, diffusion policies, VLAs, world models, sim-to-real, active learning, TensorRT, ONNX, Triton, CUDA, Isaac Sim, MuJoCo, or ManiSkill is valuable. Robotics experience is excellent, but exceptional candidates from vision, video, autonomous systems, or large-scale generative modeling should also apply.