# Founding ML / RL Engineer — Silicon Dreams

- Location: Menlo Park, CA
- Posted pay: $125k–$250k/yr
- Posted: Sep 28, 2026
- Status: open (checked Nov 20, 2026)
- Apply (employer's site): https://jobs.ashbyhq.com/silicondreams/b8898e0b-50ae-4d2b-9d78-816aeee56efb
- Employer record: https://visajobs.io/company/silicon-dreams.md

## Sponsorship history
- Tier: Employer history only
- With this exact title: 0

## Description
About Silicon Dreams
Silicon Dreams is ushering in a future of true abundance for humanity. We are fundamentally rebuilding the way that production and manufacturing is done from the ground up.
Raw material in.
Electricity in.
Human desire in.
Matter out.
We are focused on solving real problems for humans. We believe the most important problem to solve is accelerating a future of abundance across all goods that elevate human quality of life.
About the Role
We are looking for an ML / RL Engineer to own the learning stack that turns simulation and real-world data into robot policies that work in a factory — from the training objective and the architecture, to the distributed training runs, to the deployed policy running on real hardware at rate. You will be the extreme owner of learning from dataset to deployment, from the first overfit-on-one-scene sanity check to a policy running thousands of production cycles.
This is a generalist role by design. In a given week you might design a reward and find out it is being gamed, rebuild a data pipeline because the state representation was wrong, train an imitation policy on teleoperated demonstrations, run an RL fine-tune across a GPU cluster, and then go out to the bench to find out why the policy that was flawless in sim stalls on contact.
You will have creative freedom to architect from first principles, but you will also own the unglamorous parts: dataset hygiene, evaluation protocols that cannot be gamed, checkpoint provenance, and the release discipline that keeps a regressed policy off real hardware.
We care about policies that work, not benchmark numbers. A result counts when it holds across randomized conditions on real hardware, with the failure modes named and bounded — not when a success rate looks good on the scenes it was tuned on.
You need to be a team player, owning all aspects of the end outcomes, as learning here lives at the seam between simulation, perception, controls, and the physical machine, and the hard problems are almost always cross-domain.
What You'll Do
Design, train, and ship robot policies — imitation learning, reinforcement learning, and the hybrids that actually work on contact-rich manipulation
Own the data path end to end: collection from teleoperation and from simulation, labeling, state and action representations, augmentation, and the dataset versioning that makes a run reproducible
Build and own the training infrastructure — distributed training across multiple GPUs and nodes, experiment tracking, hyperparameter sweeps, and cost per experiment
Own evaluation: held-out scenes, randomized conditions, real-hardware trials, and metrics that measure the task rather than the training distribution
Close the sim-to-real loop — domain randomization, calibration against measured behavior, and identifying which fidelity gaps actually change the policy
Deploy policies to real robots: inference latency budgets, on-device or edge execution, safety envelopes, and graceful fallback when the policy is out of distribution
Instrument everything for observability — rollout logging, failure capture, and diagnostics that make it possible to debug a bad episode after the fact
Own model release and traceability: what checkpoint is running on which unit, trained on what data, evaluated how
Collaborate closely with the simulation, perception, and firmware work, co-defining the interfaces where policy performance is coupled to what the rest of the stack provides
What You'll Bring
2+ years of experience (engineering club, commercial and research internships included) training and shipping machine learning models, ideally for control or robotics
Strong proficiency in Python and PyTorch (or JAX), with a feel for where training time, memory, and throughput actually go
A real grounding in reinforcement learning and imitation learning — policy gradients, actor-critic methods, behavior cloning, and an honest sense of when each one fails
Experience building data pipelines for learning: you have been burned by a dataset bug and now check the data before blaming the model
Comfort in a Linux environment and on a job scheduler (Slurm or similar), running many experiments and accounting for all of them
Strong debugging skill across the model/system boundary: you can tell whether a bad rollout came from the reward, the data, the architecture, the simulator, or the hardware, and you design the experiment that decides it
Strong first-principles problem-solving and a focus on maintainable, reproducible work — you care that a result holds on hardware you have not tuned against, not just on the eval split
Ability to work independently in a fast-moving, cross-functional environment
If you don’t meet every requirement listed above, but believe you are exceptional, believe in our mission, and are willing to work incredibly hard, please apply regardless. We always want to work with the best.
Preferred Skills
Bachelor's or Master's degree in Computer Science, Electrical Engineering, Robotics, Applied Mathematics, or related
Experience with contact-rich manipulation, dexterous control, or mobile manipulation on real hardware
Experience with modern robot-learning architectures — diffusion policies, action chunking, transformer or vision-language-action models
Experience with model-based RL, world models, or learned dynamics for planning
Experience with teleoperation systems and demonstration collection at scale
Sim-to-real transfer: domain randomization, system identification, and knowing what the gap actually costs
Classical controls background — MPC, trajectory optimization, impedance and force control — and comfort mixing it with learned components
Large-scale training experience: multi-node, mixed precision, sharding, and data loading that keeps the GPUs fed
Model compression, quantization, or inference optimization for real-time on-robot execution
Familiarity with Git, CI/CD, and reproducible research workflows
Background in robotics, mechatronics, or ph

> History, not a promise: past H-1B filings do not mean an employer will sponsor you or a given role. Data: U.S. DOL OFLC LCA disclosure (FY2024–FY2026 for role counts, certified LCAs) and the USCIS H-1B Employer Data Hub. Not legal advice. Methodology: https://visajobs.io/methodology
