XDOF is trying to own the layer under the current robotics boom: the demonstration data. Rather than train its own robot policy, it combines remote teleoperation with human collectors who wear sensors to record everyday tasks like folding clothes and flattening boxes, then sells those recordings as training data. Per TechCrunch’s September 4, 2026 report, the company was founded in 2024 by UC Berkeley researchers Philipp Wu (CEO) and Fred Shentu (CTO), came out of stealth in June 2026 with a $70 million Series A backed by Thrive Capital, Andreessen Horowitz, Lux, and Spark Capital, and already serves around 20 customers, several of them frontier AI labs. That same report said XDOF was in late-stage talks for a Series B near a $1.2 billion valuation led by 8VC, with terms that could change and no close announced, so treat the $1.2B as a reported target, not a done deal.
XDOF is not a robot-foundation-model shop, the kind our Skild AI, Physical Intelligence, and Generalist guides cover. Those companies build the policies; XDOF builds and sells the data that policies are trained on. So the interview leans harder on data quality, pipeline throughput, and the mechanics of capturing clean multimodal recordings than on inventing a new learning algorithm. If you are weighing several private-company bets in this space, our AI-startup interview guides help you sort the data-layer companies from the model labs.
Why the questions look the way they do
A data-supply company lives or dies on whether the demonstrations it ships actually make a robot better. That single fact shapes the whole loop, and since this is inferred from the business model, not a reported account, read it as the logic behind the questions, not a transcript. A teleoperation rig that produces subtly corrupted episodes, misaligned timestamps, a drifting camera, a gripper-state channel that lags the video, is worse than useless: it silently degrades a customer’s training run. So expect interviewers to keep pulling on data integrity: how you would catch a bad demonstration before it ships, how you know the sensor streams are actually synchronized, how you measure whether a batch of episodes improved a downstream policy at all. Candidates who talk only about model architecture, never the data feeding it, are what this loop screens out.
What the loop probably contains
No candidate has posted an XDOF interview report as of September 2026; the company is only a few months into public existence, with reporting around the Series B talks putting the team near 60 people. So treat the stages below as a representative robotics-and-data loop reconstructed from how comparable teleoperation, robot-learning, and data-infrastructure companies interview, not as XDOF’s published process. The round names are this guide’s labels, not XDOF’s terminology, and the durations are peer-typical rather than anything XDOF publishes. At roughly 60 people a founder conversation is near-certain, so the table includes one; some peers at this stage swap the live track screen for a take-home. Confirm the actual shape with your recruiter on the first call.
| Representative XDOF interview stage (reconstructed, not reported) | Format and length | What it screens for |
|---|---|---|
| Recruiter screen | ~30 min phone | Background, which track you fit (robot learning, data infrastructure, teleop systems), work authorization, location and on-site expectations |
| Hiring-manager technical call | 45-60 min | A data or robotics system you owned end to end: what you built, what broke, how you measured quality |
| Track-specific screen | 60 min | Live coding plus domain depth: imitation learning (robot learning), pipeline and storage design (data infra), or real-time systems (teleop) |
| On-site: coding round | ~60 min | Data-structure and algorithm work close to the job, often stream processing or dataset manipulation, not abstract puzzles |
| On-site: system or ML design | ~60 min | Petabyte-scale data pipeline, a training-data quality system, or a low-latency teleoperation stack, depending on track |
| On-site: cross-boundary round | ~45-60 min | How your part connects to the rest: collection hardware, annotation, storage, and what the customer’s training run needs |
| Founder round (CEO Wu or CTO Shentu) | ~45 min | Near-certain at this size: technical taste, how you reason about data quality, and conviction that the product is the data, not the robot |
| Behavioral / values round | ~45 min | Ownership, moving fast in an early data-operations org, working across engineering and the field collection teams |
Interviewers keep drilling into quantitative follow-ups until you run out of answers: name a throughput figure and they ask what bounded it, claim a data-quality gain and they ask how you measured it. Bring the real numbers, including the ones that disappointed you.
Robot learning and ML roles
If you interview on the learning side, the questions center on what makes demonstration data good enough to train on. Representative prompts: how you would detect low-quality or inconsistent demonstrations in a large teleoperation dataset before they poison a policy; behavior cloning and where it fails, covariate shift and compounding error, and what DAgger or similar approaches buy you; how much data a given manipulation skill actually needs; and why a policy trained on one robot embodiment transfers poorly to another. Expect an evaluation question too, because a data company has to prove its product works: how would you show a new batch of episodes improved a customer’s success rate, not merely enlarged the dataset. These are standard robot-learning problems, not XDOF secrets.
Data infrastructure roles
The data-supply thesis is really an infrastructure bet, so this is likely the heaviest hiring track. Expect to design at scale: a pipeline that ingests synchronized multi-camera video, robot proprioception, and wearable-sensor streams from collection sites around the world, validates and time-aligns them, and lands them as training-ready episodes. Representative ground includes petabyte-scale storage and retrieval built for training runs that read the same data many times, deduplication and lineage so a customer can trace exactly which episodes went into a model, and schema design for heterogeneous multimodal records. A good design round wants a real tradeoff, not a component list: when collection outruns processing, do you apply backpressure and slow the collectors, or drop-and-backfill so field sites never stall and reconcile the gaps later. Backpressure protects completeness but wastes paid collector time; drop-and-backfill keeps sites productive but risks losing an episode you can never recapture, so the right call depends on how repeatable the task is. The system-design interview guides cover the pipeline patterns; the separate coding round leans on data-structure and algorithm work, where a refresh on time and space complexity and the common coding patterns helps.
Teleoperation and robotics systems roles
Someone has to build the rigs and software that let an operator drive a robot smoothly enough to produce useful demonstrations, and that is its own discipline. Expect real-time systems questions: the end-to-end latency budget for a teleoperation loop and what degrades first when the network gets bad, how you keep control stable under jitter and packet loss, and how you time-sync camera frames against control commands so the recorded episode reflects what happened. Representative prompts: design the capture path so every sensor stream carries a trustworthy timestamp; handle a dropped video frame or a lagging force channel mid-episode without corrupting the recording; and instrument a fleet of collection stations so a degrading rig announces itself before it ships bad data.
The collection and operations track
XDOF also hires and trains teams of human data collectors and teleoperators, a separate, largely non-engineering track with its own screening. If you are applying to build the systems rather than operate them, clarify with your recruiter which pipeline you are in.
Work authorization and location
XDOF’s careers page listed three offices in September 2026: San Mateo, California, plus Mexico and Jakarta, with roles tagged hybrid or on-site and none remote, so expect real in-person presence wherever the req sits. San Mateo engineering roles skew hybrid, while the Mexico and Jakarta postings, weighted toward collection operations, are on-site. What the page does not publish is visa sponsorship, so ask the recruiter directly whether your exact req sponsors.
How to read the comp
No public XDOF salary range was reachable while writing this, so anchor to comparable roles elsewhere until you see the offer. A levels.fyi query for robotics and ML engineers at well-funded startups puts base around $160,000 for a mid-level engineer up past $250,000 for staff, before equity; run that same query for your level to get the current spread, and expect an early company to weight the package toward stock. If your req shows a pay-transparency range in a state that requires one, that is the most reliable XDOF-specific number, so read it there. On the equity, be careful: the $1.2 billion figure is a reported Series B target that had not closed as of this writing, and no Series A post-money valuation was published, so there is no clean number to value a grant against. Ask the recruiter for the last preferred price per share, the current 409A valuation, and the fully diluted share count, then pin down whether it is options or RSUs, the strike or grant price, the vesting schedule, and your level. If the roughly $1.2B Series B closes while you are in process, ask when the 409A was last reset, since it may reprice the grant. Run the numbers through a total-comp calculator once you have real inputs, and a salary-negotiation walkthrough covers how to push back. A grant this early might pay off big or come to nothing, so count it as upside rather than salary.
How to prepare
Prepare by track, not by grinding generic problems. Robot-learning candidates should talk about imitation learning and data quality with equal fluency, since this company cares more about the data than a research lab does. Data-infrastructure candidates should rehearse a full multimodal ingestion-and-storage design with real throughput and cost numbers. Teleop candidates should have a latency budget and a synchronization story ready. Everyone gets a project deep-dive, so choose one or two efforts you owned and be ready to keep answering past the headline, down to the measurement that caught you off guard and what you did; if your stories wander, the STAR structure keeps them tight. Spread the prep over a couple of weeks with a lightweight study plan, and run your resume through a resume checker so it quantifies what you shipped.
Two things decide this loop. Can you own your part down to the measured numbers, and do you understand that the product here is the data, not the robot. Nail the first, miss the second, and the cross-boundary round is where they make the call.
Practice the behavioral round:
