Skild AI will spend more of your interview asking why your last policy fell over on real hardware than asking you to reverse a linked list. That one fact tells you what kind of company this is. The Pittsburgh startup, founded in 2023 by former Carnegie Mellon professors Deepak Pathak and Abhinav Gupta, is trying to build a single model, the Skild Brain, that controls any robot for any task without retraining per platform. In January 2026 it closed a $1.4 billion round led by SoftBank at a valuation above $14 billion, with Nvidia’s NVentures, Bezos Expeditions, Samsung, and LG also in. Raise that kind of money on a research bet and the interview mirrors the bet. They want people who can drag an idea out of a paper and onto a robot that survives contact with a real floor.
Who they hire, and why the tracks feel different
Skild splits roughly into four hiring tracks. Research scientists publish on generalist policies and robot learning. Research engineers turn those ideas into training runs that don’t quietly diverge overnight. ML and data infrastructure people build the pipelines that pull teleoperation logs, simulation rollouts, and raw human video into one training corpus. Platform and full-stack engineers build the cloud product that robot makers call through an API. The loop you get depends on which of these you’re chasing, so pick one before you prep. A research candidate gets pressed on generalization and sim-to-real. A backend candidate gets pressed on throughput and failure modes.
What holds across every track is the taste bar. Pathak and Gupta come out of academia, where the only currency that counts is whether the thing works and whether you can say precisely why. Hand-wave and you’re done. Claim domain randomization closed your sim-to-real gap and you’ll get an immediate follow-up: which parameters did you randomize, over what ranges, and how did you rule out that you just overfit the evaluation.
The loop you’ll actually face
Skild doesn’t publish its process, and it’s young enough that the steps still shift from month to month and team to team. What candidates report, and what’s standard for a lab with this pedigree, tracks closely to the table below. Treat it as the shape, not a contract.
| Stage | Format | What they are testing |
|---|---|---|
| Recruiter screen | About 30 min, call | Why robotics, why Skild, and which track your background fits |
| Technical phone screen | 45-60 min | Coding fluency for infra roles, or depth in RL and imitation learning for research roles |
| Problem session | 60 min, shared doc or whiteboard | Reasoning out loud through an open robot-learning or scaling problem |
| Coding round | 45-60 min | Data structures, vectorized numerical code, and whether your code is readable |
| Project talk | 45-60 min | Presenting your own work and defending every design choice under questioning |
| Design or infra round | 60 min | Training pipelines, data at robot scale, or the robot-facing cloud API |
| Team and leadership chats | 30-45 min each | Ownership, comfort with ambiguity, and whether you actually want to work on physical AI |
The research round: robot learning, not LeetCode
If you’re going for a research or research-engineering seat, this is where the offer is won or lost. The questions start from real systems and push on the parts most people gloss over. Expect prompts phrased like these:
- “You have 10,000 hours of teleoperation data and 100 hours of real robot rollouts. How do you combine them to train one policy?”
- “Behavior cloning compounds errors at test time. Explain why, and explain what DAgger actually fixes and what it doesn’t.”
- “Train a quadruped to walk on ice it never saw in simulation. Where does sim-to-real break first, and what do you change?”
- “When would you learn a world model instead of a policy directly, and what does that buy you?”
- “Your reward is sparse and the robot never stumbles into success. Walk me through your options before you reach for reward shaping.”
The people running these rounds have read the same papers you have, so name-dropping RT-2 or Open X-Embodiment won’t carry you. What lands is picking one generalist-policy result, saying what it got right, and saying where you think it’s brittle. If you’ve trained anything on a real robot, lead with the failure. The gap between “it worked in sim” and “it worked on the arm in the lab” is the exact territory Skild is trying to own, and candidates who’ve lived in that gap read as credible in about ninety seconds.
The systems round: training and data at robot scale
For infra and research-engineering roles, the hard questions are about moving data and keeping big training runs alive. A generalist policy is only as good as the corpus behind it, and that corpus is a mess of formats: teleoperation from many robot types, physics-sim rollouts, and raw human video. Expect to design the pipeline that ingests all of it, aligns timestamps and action spaces, and serves batches fast enough that your expensive GPUs aren’t sitting idle.
Concrete versions of this: how would you run several thousand parallel simulation environments and find the throughput bottleneck when your GPUs are starved? How do you shard a large policy across accelerators, and what breaks first as you scale the batch size? How do you make a training run resumable when a node dies six hours in? They care less about the buzzword and more about whether you’ve felt these problems. If you’ve debugged a data loader that turned out to be the real bottleneck behind a “slow model,” that story is worth more than a tidy diagram.
The coding screen
The coding round is standard data structures and algorithms with a numerical tilt. You’ll see array and graph problems, and often something that rewards clean vectorized thinking over a triple-nested loop. A representative warm-up: given the rewards from one episode and a discount factor, return the discounted return at every timestep in a single pass.
def returns(rewards, gamma):
out = [0.0] * len(rewards)
running = 0.0
for t in reversed(range(len(rewards))):
running = rewards[t] + gamma * running
out[t] = running
return out
Nothing exotic, but they watch how you get there. Talk through the reverse pass, state the time and space complexity without being asked, and write code a teammate could read. Comfort with numpy or torch tensor operations helps more than memorized tree traversals, since that is the code you would actually write on the job.
Comp, and what the range depends on
Skild is private and pre-IPO, so total compensation leans hard on equity, and the equity is the part worth negotiating. Cash bases for senior engineers at frontier AI companies generally sit in the band you would expect from top Bay Area and remote labs, with total comp for strong senior and staff candidates commonly landing in the mid-six figures once equity is counted. At a company that tripled its valuation in seven months, the spread between a good and a great equity grant dwarfs the base difference. Numbers move quickly at this stage, so treat any figure you read as stale and check Levels.fyi and your recruiter for the current band on your specific level and location rather than trusting a blog post.
One thing to weigh: equity in a $14 billion private company is not lottery-ticket cheap anymore, and the upside is more capped than it was at the earlier $1.5 billion mark. You are trading some of that ceiling for a much higher chance the bet pays off at all. Whether that trade is good depends on your risk tolerance and what stage of your career you are in.
How to prep without wasting a month
Read the Skild Brain writeups on the company’s site and skim the founders’ CMU work, since it signals what they value: Pathak’s curiosity-driven exploration and rapid motor adaptation, Gupta’s self-supervised learning. Be ready to talk through one recent generalist-policy paper and say where you think it’s wrong; a clean summary won’t carry the round. Rehearse your own project as a story that leads with the hard part and the failure, because the project talk is where mid candidates and strong ones separate. For infra roles, get fluent in parallel simulation, distributed training, and the un-glamorous reality of data pipelines. For coding, keep your data structures sharp and practice writing vectorized numerical code by hand.
The candidates who do well here aren’t the ones with the spotless competitive-programming record. They are the ones who have had a robot fail on them in the physical world and can tell you, without flinching, exactly why it happened.
