Machine Learning Engineer, Robotic Foundation Models

About PhyAgents

PhyAgents is a high-growth foundation model startup headquartered in San Jose, California. Founded by Stanford alumni, it sits at the intersection of the two biggest platform shifts of our time: foundation models and robotics. PhyAgents is backed by three Stanford professors, top AI leaders, and deep tech investors.

PhyAgents is building the Claude Code for Physical AI: a robotic foundation model plus a physical agentic framework for reliable, long-horizon operations in industrial environments. It delivers semi-humanoid robots as a workforce for industrial applications.

Their “physical agents” (or robots) combine a robotic foundation model, force-aware manipulation, long-horizon reasoning, step-by-step execution verified against real operating procedures, and human-in-the-loop autonomy. The robot plans the job, executes each step, checks their work, and when it isn’t certain it doesn't guess—it escalates.

About The Team

• CEO and co-founder: Serial entrepreneur and former VP at Intel. Previously raised >$50M, scaled and exited a robotics company. PhD from Stanford.
• CTO and co-founder: Key foundation models leader from Apple. Previously led perception for self-driving. MS from Stanford, B.Tech from IIT Madras.

The team includes executives and AI leaders from Apple, Intel, Marvell, Samsung, Uber, Panasonic, VMWare and Lockheed Martin.

Why join now

You’ll be among the first engineers in the India team of a Silicon Valley deep tech startup. You’ll work on the same problems, with the same bar, as our team in San Jose. This is not a satellite office working off tickets. The India team is a core part of how we build models, data, and infrastructure, and you’ll own real pieces of the stack from your first quarter.

Here you’ll own significant components, work across the whole stack, contribute at the cutting edge, see your code move a robot, and grow as fast as the company does.

The physical AI field has shown broad competence across many tasks, and almost never 99.9% reliability on even one. That gap is the company. Closing it takes models that learn from demonstration and improve through practice, a data flywheel that turns every failure into training signal, and software that runs on real robots in real industrial environments.

About The Role

At PhyAgents, a machine learning engineer works across the full loop that makes a robot better: data, models, training, simulation, evaluation, and deployment on real hardware. You’ll do research-grade work and ship it onto robots in the same quarter. A result that doesn’t run on the robot isn’t finished.

What You Will Work On

• Training and post-training: Train and post-train robotic foundation models, including vision-language-action models, using imitation learning and reinforcement learning. You’ll own the work from data loading through multi-GPU training runs.
• Training infrastructure: Build and scale distributed multi-node GPU training, data loaders that keep GPUs fed, and reproducible pipelines with experiment tracking.
• Simulation: Build simulation environments for our tasks. Train policies with reinforcement learning that transfer to real robots.
• Rewards and verifiers: Design the reward functions and the verifiers that check whether a task step actually succeeded.
• Evaluation: Build harnesses that tell us whether a model actually got better, in simulation and on real hardware. Design the metrics that decide what ships.
• Data and data infrastructure: Define what data the models need. Build the pipelines that turn raw robot episodes into versioned, training-ready datasets. Trace a stuck policy back to its data.
• Models on robots: Get models running reliably and fast enough on real hardware. Debug failures in the lab across the model, the data, and the robot.

What We’re Looking For

Minimum

• 5+ years of hands-on experience training and shipping ML models in production or in a research lab, with a B.Tech, M.Tech or PhD in Computer Science, Electrical Engineering or a related field
• A track record of owning an ML system end to end: you’ve taken something from problem definition through data, training, evaluation, and deployment, and you can tell us what you’d do differently
• Deep hands-on experience with PyTorch, including writing and debugging non-trivial training loops
• Experience training at scale: multi-GPU or multi-node training (DDP, FSDP, DeepSpeed, or equivalent) and the profiling work to keep it efficient
• Strong fundamentals in linear algebra, probability, optimization, and machine learning, and in data structures, algorithms, and systems
• Strong software engineering habits: you write code others can build on, and hold a high bar for code quality.

Preferred

• Experience post-training transformers, diffusion models, or multimodal models, especially with imitation learning or reinforcement learning
• Experience in robotics, autonomous vehicles, or another domain where models act in the physical world and failures are expensive
• Experience with robot simulation (MuJoCo, Isaac Sim, PyBullet) and sim-to-real transfer
• Experience building data pipelines for large multimodal datasets (object storage, Ray or Spark, Parquet, WebDataset)
• Experience deploying models to constrained or real-time hardware (ONNX, TensorRT, quantization, latency work)
• Publications at top AI or robotics venues, or meaningful open-source contributions
• Experience mentoring engineers or leading a small technical effort

What We Value

• Ownership. You finish things. A result that doesn’t run on the robot isn’t finished.
• Rigor. You catch the subtle bug that the dashboard says is fine.
• Range. You’re as willing to debug a data pipeline or a ROS node as a training run.
• Comfort with real hardware. Our models run on robots, and you’ll spend time in the lab.
• Speed of learning. There’s no playbook. You’ll be learning new parts of the stack every month.
• Ambition. You want to build something that matters at the frontier of AI and robotics.
• Clear communication. You’ll work daily across San Jose, Bengaluru, and our data operations team in Nepal.