AI Researcher

1X Technologies AS•San Carlos, CA

17h•Onsite

About The Position

We’re building humanoid robots that work in home - doing the chores, handling the tasks, and giving people their time back. NEO is our flagship - a home robot designed to move, learn, and operate in the real world alongside real people. We’re not demoing it - we’re shipping it. The 1X World Model Lab is an embodied AI research organization focused on pretraining the foundation models to accelerate the emergence of embodied intelligence. As the lab grows, researchers contribute where they have the most leverage, and the problems worth solving span every layer of the stack. The lab is founded on a simple thesis: robotics is not a fine-tuning problem. To build truly general humanoids, we need to pretrain on the most important data from the very beginning.

Requirements

Strong Python and PyTorch (or equivalent deep learning framework), with experience in large-scale codebases and ML tooling
Demonstrated experience in at least one area of the ML stack: large-scale model training, data pipeline engineering, ML infrastructure, or systematic model evaluation
Degree in Computer Science, Machine Learning, or a related field; graduate-level education or equivalent research experience strongly preferred
Track record of impact: published research, deployed production ML systems, or infrastructure that measurably accelerated a team's work
Research depth plus engineering rigor conducting frontier research and builds systems others depend on; doesn't treat production engineering as someone else's job, and pushes work past promising training curves to deployed capabilities.
Full-stack ML thinker understanding the path from raw robot data to trained model to deployed policy, and can identify and address bottlenecks at any layer of that stack: data quality, training efficiency, model architecture, or inference performance.
Scale-first mindset believing scale is foundational to capable humanoid robotics; designs systems with 10x and 100x growth in mind, and actively pushes to remove whatever is currently the binding constraint on model improvement.
Fast, independent contributor picking up new domains and codebases quickly, identifies the highest-leverage contribution, and makes meaningful progress without waiting for a detailed spec.

Nice To Haves

Experience with distributed training frameworks (TorchTitan, DeepSpeed, FSDP/ZeRO) and/or large-scale data pipeline and ETL systems spanning on-device, on-premise, and cloud infrastructure
Experience with multi-modal generative models, RL policy training, or sim-to-real transfer; familiarity with world models, diffusion models, or autoregressive architectures for continuous observations and actions
Experience with inference optimization techniques: quantization (PTQ, QAT, INT8/FP8), CUDA/Triton kernel development, or serving systems (TensorRT or equivalent)
Background in embodied AI, robotics, or physical systems where model predictions must translate reliably to real-world actions

Responsibilities

Advance NEO's intelligence by building the AI systems, infrastructure, and data engines that enable the robot to learn from experience and become increasingly capable in real-world environments.
Build large multi-modal generative world models and RL policies that learn from robot experience, spanning model architecture, data pipeline engineering, tokenization, and training at scale.
Advance the robot's ability to predict, plan, and act in unstructured environments.
Design and operate the data engine that turns fleet experience into training-ready datasets: intelligent upload triggers, ETL pipelines, annotation interfaces, automated labeling, and the tooling that makes robot data queryable and useful at scale.
Own the distributed training and inference systems that keep compute fully utilized. Think GPU training runs, fault-tolerant experiment management, inference optimization (quantization, kernel engineering, distillation), and on-device policy deployment.
Build the evaluation infrastructure that connects pre-training metrics to real-world robot performance: benchmarks, evals frameworks, model ranking systems, and the tooling that lets the team iterate on architectures with confidence that lab results predict what happens in the field.
Advance robot capabilities through research, improving model architectures, scaling data pipelines, optimizing training or inference, or building evaluations that make lab results predictive of field performance.
Build infrastructure that multiplies team research velocity: pipelines that are faster, evaluations that are more predictive, training systems that are more efficient, or tooling that eliminates manual work across the lab.
Ship research to production: own the path from experimental result to deploy capability on robot hardware, and measure impact by what NEO can do, not just what the model achieves on benchmarks.
Contribute to a learning flywheel where more robot experience leads to better models, better models enable more capable robots, and more capable robots generate richer experience.