AI Engineer (World Models)

Socket.dev; San Francisco (CA); Hybrid; USD 190,000 - 230,000; Full time

Job summary

Socket.dev in San Francisco is seeking a hands-on engineer to architect world models for a general-purpose humanoid robot. You will design model-based RL architectures, implement predictive world simulators, and bridge research prototypes to production-ready components.

You should be fluent in Python with production experience in C++, work with ROS2 and robotics simulators (Isaac Sim, MuJoCo), and read recent arXiv papers to push the state of the art.

Qualifications

• Hands-on experience building world models, model-based RL, or predictive world simulators using frameworks like PyTorch or JAX — you have shipped these systems, not just studied them
• Strong foundation in deep learning architectures relevant to world modeling: transformers, diffusion models, neural radiance fields (NeRF), and variational recurrent state-space models
• Proficient in Python as a primary research and development language, with production-level familiarity in C++ for latency-sensitive inference and real-time robotics integration
• Experience with robotics middleware and simulation environments — ROS2, Isaac Sim, MuJoCo — and the ability to close the sim-to-real gap in learned representations
• Experience with video prediction or future-frame generation models (e.g., RSSM, DreamerV3, UniSim, Genie) is a strong plus
• Able to read and implement from recent arXiv papers with minimal overhead — you are comfortable turning a research prototype into a tested, integrated system

Responsibilities

• Architect and implement world models and model-based RL components for the humanoid platform
• Bridge research prototypes to production-ready modules with robust interfaces
• Collaborate with hardware, perception, and control teams to integrate learned representations
• Evaluate and iterate on simulation-to-real transfer using established robotics simulators

Skills

World models

Model-based RL

Python

C++ robotics

ROS2

Simulation environments

Video prediction models

ArXiv reading

Tools

PyTorch

JAX

ROS2

MuJoCo

Isaac Sim

UniSim

DreamerV3

Job description

Why are We Hiring for this Role:

• We are building a general-purpose humanoid that must understand and navigate the physical world — and that requires a dedicated engineer to architect the internal models that make that possible
• World models are the cognitive backbone of our robot; without them, the humanoid cannot plan, predict, or adapt to novel environments
• We are at an inflection point where our hardware is ready — now we need the intelligence layer to match it
• The gap between a robot that executes fixed commands and one that truly reasons about its environment is a world model; we are hiring to close that gap
• As we scale to real-world deployment, our humanoid needs to generalize across unstructured, unpredictable settings — something only a robust world model can enable
• This hire will directly shape the core intelligence architecture of our platform before it becomes locked in at scale

What Kind of person are we looking for

• Hands-on experience building world models, model-based RL, or predictive world simulators using frameworks like PyTorch or JAX — you have shipped these systems, not just studied them
• Strong foundation in deep learning architectures relevant to world modeling: transformers, diffusion models, neural radiance fields (NeRF), and variational recurrent state-space models
• Proficient in Python as a primary research and development language, with production-level familiarity in C++ for latency-sensitive inference and real-time robotics integration
• Experience with robotics middleware and simulation environments — ROS2, Isaac Sim, MuJoCo — and the ability to close the sim-to-real gap in learned representations
• Experience with video prediction or future-frame generation models (e.g., RSSM, DreamerV3, UniSim, Genie) is a strong plus
• Able to read and implement from recent arXiv papers with minimal overhead — you are comfortable turning a research prototype into a tested, integrated syste