Research Scientist

We combine world-class research with top-notch engineering and apply it to solve real problems.

Much of this data already exists. We have robots in production at scale. We aren't waiting for datasets to be collected; the byproduct of our machines doing useful work is being created right now.

We measure what matters. We test our code in unit tests, simulations, and directly on real robots. Grounding our models in deployment allows us to truly measure performance, not just offline metrics.

High leverage, high impact. We’re still a highly focused team. If your architectures and training curricula improve our agents, you directly change the economics of the company.

World-class peers. Our team has built Google Warsaw, unicorn startups, led research in DeepMind, tested rocket engines, and worked at top companies like Nvidia and ByteDance. Now, we are shaping the reality of Physical AI together.

We are building the bridge. We aren't a new startup looking for an application; we are an established player bootstrapping physical AI. We believe this will be the first true proof-of-concept for scaled physical AI

Your focus will be defined by the intersection of ML research, robotics, and large-scale multimodal model training. Expect challenges across two main pillars, with the opportunity to specialize in Pretraining or Post-training:

Foundation Models & Pretraining

Design the Base Intelligence: Define model architectures (Transformer- and Diffusion-based), objectives, and training curricula across multimodal robotic data, turning raw deployment logs into generalizable capabilities.

Master the Data: Develop scalable data mixtures and sampling strategies utilizing our massive offline repositories of vision, action, and state data.

Push the Frontier: Run rigorous ablations to understand scaling laws, data quality effects, optimization dynamics, and large-model failure modes.

Scale with Engineering: Collaborate closely with ML Infra to push cluster utilization and throughput, ensuring our algorithmic ideas translate to efficient distributed training.

Adaptation, Post-Training & Real-World Evaluation

Drive Downstream Adaptation: Explore fine-tuning recipes to make general models – our own as well as our partner’s models – useful, controllable, and safe in the real world using imitation and reinforcement learning, distillation, and curriculum learning.

Improve Physical Robustness: Develop cutting-edge methods for improving real-world reliability, handling out-of-distribution edge cases, and steering robot behavior in mature factory environments.

Build Benchmarks: Design evaluation frameworks and lightweight physical setups that measure actual robot performance and failure modes far beyond the limits of simulation.

Close the Physical Loop: Analyze real-world evaluation results to guide the overarching research direction, seamlessly bridging the gap between foundation model outputs and physical-world outcomes.

Experience: Deep research and practical experience at the intersection of machine learning, systems engineering, and physical robotics.

Proven Track Record: Experience designing, training, and fine-tuning large-scale deep learning architectures (VLMs, VLAs, RL, RLHF, Imitation Learning), ideally with policies deployed and validated on real hardware.

Engineering Excellence: Strong deep learning framework fundamentals (PyTorch/JAX). You are comfortable debugging at every layer of the stack and care about empirical rigor as much as raw iteration speed.

Robotics Intuition: Comfort working hands-on with hardware. You understand the robotics full stack (perception, controls, state estimation) and care deeply about evaluation and failure analysis when software meets the physical world.

Pragmatic Research Mindset: You possess the ability to move seamlessly between theoretical design and physical implementation. You prefer execution, rapid iteration loops, and real-world robustness over academic purity.