Robotics Research Engineer

Frontier Arc; San Francisco (CA); On-site; USD 180,000 - 260,000; Full time

Job summary

Frontier Arc in San Francisco seeks a Research Engineer to push the core RL research agenda. You’ll own the end-to-end loop—from designing policies to training, evaluation, and interpretation—working on a proprietary simulator that aims to replace costly real-world data collection.

You’ll join a small, technically serious team with high agency, building models that transfer from sim to real hardware. Expect on-site work, a fast-paced lab environment, and the opportunity to tackle difficult

Qualifications

• You have designed and trained RL policies from scratch with full ownership of architecture and training.
• Hands-on experience with Isaac Lab, MuJoCo, Newton, or a close equivalent modern physics simulator.
• You understand models under the hood, with interpretability and ablation analysis to reason about failure modes.
• Track record of high-agency, end-to-end ownership of difficult technical problems.

Responsibilities

• Design and train reinforcement-learning policies from scratch, end to end: full ownership of architecture, reward structure, and training process.
• Run ablations, analyse failure modes, and iterate until production-worthy performance.
• Use model interpretability techniques to understand policy behavior and improve it.
• Own the full research loop end to end and ensure sim-to-real relevance.

Skills

RL policy design

End-to-end ownership

Research mindset

Tools

MuJoCo

Isaac Lab

Newton

Job description

As a research engineer, it's likely you've spent meaningful portions of your career waiting.

Waiting for GPU allocations. Waiting for approval to test a new approach. Waiting for a senior researcher to define the problem.

At this robotics research lab, the constraint has been removed.

What remains is the actual research, which is harder and more interesting than anything you can do at a larger organisation where you own a narrow slice of a well-defined pipeline.

The team is small and technically serious. The founder still spends most of their time training models.

The people you will work alongside have been selected for the same profile: high agency, end-to-end capability, a track record of doing impressive things in difficult problem spaces.

You'll be working on the core technical bet of the company - that a well-built simulator can replace real-world data collection as the engine of RL policy learning. You'll be inside what may turn out to be one of the most important research environments in physical AI, at the moment when the fundamental research the field has been waiting on has just become possible.

You'll get your own personal GPU cluster, unlimited compute for training runs, and a personal humanoid to test on. No shared resources, no paperwork, no waiting in line. The experiments and model ideas you couldn't run at your last lab because you were rate-limited by data are now possible.

San Francisco (on-site, 5 days a week)

As a research engineer, it's likely you've spent meaningful portions of your career waiting.

Waiting for GPU allocations. Waiting for approval to test a new approach. Waiting for a senior researcher to define the problem.

At this robotics research lab, the constraint has been removed.

What remains is the actual research, which is harder and more interesting than anything you can do at a larger organisation where you own a narrow slice of a well-defined pipeline.

The team is small and technically serious. The founder still spends most of their time training models.

The people you will work alongside have been selected for the same profile: high agency, end-to-end capability, a track record of doing impressive things in difficult problem spaces.

You'll be working on the core technical bet of the company - that a well-built simulator can replace real-world data collection as the engine of RL policy learning. You'll be inside what may turn out to be one of the most important research environments in physical AI, at the moment when the fundamental research the field has been waiting on has just become possible.

You'll get your own personal GPU cluster, unlimited compute for training runs, and a personal humanoid to test on. No shared resources, no paperwork, no waiting in line. The experiments and model ideas you couldn't run at your last lab because you were rate-limited by data are now possible.

What you\'ll do

The day-to-day is pure model research inside the proprietary simulator:

• Design and train reinforcement-learning policies from scratch, end to end: full ownership of architecture, reward structure, and training process.

• Run ablations, analyse failure modes, and iterate on the model until success rates are genuinely production-worthy. The bar is 95%+ on target tasks.

• Use model interpretability techniques (attention and activation analysis, ablation studies) to understand why a policy behaves the way it does and improve it.

The sim works. Policies trained in simulation transfer to real hardware with results on par with or better than real-world trained models, which means your focus stays where it matters: the model, not the deployment pipeline.

You won't be writing reward functions in isolation or fine-tuning existing public models. You'll own the full research loop end to end.

What you\'ll need

• You have designed and trained RL policies from scratch. Full ownership of architecture, reward structure, and training process. Not fine-tuning existing models or implementing published papers with minor modifications.

• Hands-on experience with Isaac Lab, MuJoCo, Newton, or a close equivalent modern physics simulator.

• You understand how models work under the hood. Comfortable with model interpretability, ablation studies, attention and activation analysis, and reasoning carefully about failure modes.

• You can demonstrate a track record of high-agency, end-to-end ownership of a difficult technical problem - whether inside a research lab, an industry role, or independent work.

About the company

Robotics has been bottlenecked by data. Most of the field still scales by paying humans to teleoperate, which caps every research idea at what two weeks of operator time can buy. This company has solved that, building a simulator that trains zero-shot policies end-to-end in sim and deploys them to the real world. For the first time in robotics, data scales with compute rather than with man-hours.

They're a small, seed-funded research lab in San Francisco, deliberately kept lean and built around a single profile: high-agency engineers who own problems end to end. The work is pre-commercial and genuinely open-ended. There\'s no defined roadmap to inherit and no well-scoped project waiting on day one, which is exactly what makes the research worth doing.

If that sounds like the environment you\'ve been waiting for, hit apply.