Robotics Machine Learning Engineer

About Exclaim Robotics
We're a small robotics company building robots to maintain critical infrastructure, starting with data centers.
Compute density in AI data centers is skyrocketing, which is forcing changes to rack dimensions and cooling, and above all to power. Those changes make a rack a worse place for a person to work. Most of the data center is controlled remotely anyway, so the physical tasks that remain are small and very precise: replacing an optical network connector without dislodging its neighbors, reconfiguring copper wiring, and replacing NVMe drives exactly straight, all through a mess of cables.

The robot has to be reliable enough that people trust it near a rack that costs millions of dollars.
We're in pre-seed, and everything is still being built: the team, the company, the robots. If that sounds exciting rather than worrying, this might be the right place for you!

What You'd Be Doing
Our approach is to train policies in simulation first, where we know exactly where every drive bay and connector is, distill them into policies that only see what the robot can see, and then close the remaining gap on hardware using residual RL or imitation from teleoperated demonstrations. That is the current plan, and we expect whoever owns it to change it.

You'd be our first machine learning engineer, and you would own the training pipeline:

• Training infrastructure: data in, checkpoints out, with experiments tracked well enough that we can tell which change actually helped.

• Deciding what data we collect, what we keep, and what we discard. Data curation is unglamorous, and it is most of what makes a policy work.

• Building and randomizing simulation scenes that resemble a real rack closely enough to transfer.

• Taking policies onto hardware and iterating on what fails there.

You would not be doing this alone. Everyone here works on some part of the ML stack, and simulation, perception, and deployment are shared. You would be inheriting a first version of the training setup rather than starting from a blank page. Someone else will handle making the policy run fast on the robot, but we would still expect you to spend time standing next to it watching it fail, because that is where you find out what your training distribution was missing.

What You Need to Bring:

• You've trained a policy that controlled something physical, either a real robot or one in a robotics simulator. Show us a repo, a paper, or a video of it running.

• A university-level degree in robotics, machine learning, or something adjacent. If you've managed to teach yourself to fully understand research papers without one, still happy to talk.

• Depth in RL or in imitation learning: sim training and reward shaping,
  or diffusion policies, ACT, and behavior cloning from demonstrations. Both belong here, fine if you're only experienced with one.

• You can build the training pipeline yourself, data in and checkpoints out, with experiments tracked somewhere we can read them. Weights and Biases is fine, and so is whatever has replaced it by the time you read this.

• Comfortable in a simulator, Isaac Sim/Lab or Mu Jo Co  or similar, including the unglamorous half of building scenes and randomizing them.

• You don't need to be a controls or systems engineer, but you do need to care whether the policy works on hardware rather than in a notebook.

You'd be a good culture fit for us if:

• You like working closely with a small team

• You can take on a task you have no idea how to do, and figure it out as you go along

• You're comfortable giving and receiving honest feedback

• You care more about the robot working than about being right

• You're good at figuring out what needs to be done and doing it, rather than waiting to be told what to do or asking for permission

• Having an office dog isn't a deal-breaker

• You think joining an early-stage start-up will be a fun adventure :)

Don't look like a typical robotics engineer? Non-traditional background? Good. We're not really traditional either.