The thing Bedrock Robotics actually sells is a bolt-on kit: a mast of cameras, lidar, and GPS that clamps onto a stock Caterpillar excavator or dozer and lets it dig, grade, and haul across a job site with nobody in the cab. That one fact sets the tone for the whole interview. You are not building a demo that impresses a conference crowd. You are shipping software that swings forty tons of steel near human beings, and the loop is designed to find out whether you carry that weight the way the machine does.
Bedrock came out of stealth in July 2025 with $80M and a founding team pulled largely out of Waymo. CEO Boris Sofman ran Waymo’s trucking program (and founded Anki before that); CTO Kevin Peterson also comes from the self-driving world. In February 2026 they raised a $270M Series B co-led by CapitalG and the Valor Atreides AI Fund, with NVIDIA’s NVentures in the round, pushing total funding past $350M at a valuation reported around $1.75B. The company is still small enough that code you write ends up on real iron within weeks. That is the appeal, and it is also the pressure the interview probes for.
Who they’re really hiring
The org splits into a few tracks, and knowing which one you’re in changes how you should prep. There’s a perception and ML group working on 3D detection, segmentation, and depth from the sensor stack. There’s an autonomy and behavior group deciding what the machine does with that understanding. There’s an on-vehicle systems group that keeps the whole thing running in real time on hardware bouncing through a dirt pit. And there’s a data and infrastructure group feeding the models and the fleet, plus robotics application and field engineers who take all of it to a working site.
Because so many people came from autonomous vehicles, the bar reads more like an AV company than a typical Series B startup. Expect strong C++ for anything on the vehicle, Python for models and tooling, and a real interest in what happens when a sensor lies to you. The construction setting is not a softer version of self-driving. The environment is unstructured, the dust is real, GPS drops behind a spoil pile, and there’s no lane paint to lean on.
The loop
A recruiter screen comes first, and it is more technical than the average one. They want to hear that you understand why autonomy on a job site is hard and that you’ve thought about safety-critical systems, not just that you can name-drop transformers. After that comes a technical phone screen, usually coding, then an onsite (often on-site in the literal sense, at the office, sometimes with a look at the hardware) of four to six rounds.
| Stage | Format | What it screens for |
|---|---|---|
| Recruiter screen | 30 min call | Motivation, safety mindset, why robotics-in-the-real-world over a cushier lab job |
| Technical phone screen | 45-60 min, live coding | Working C++/Python, data structures, clean handling of edge cases |
| Applied coding | 60 min onsite | Practical problem solving, memory and performance awareness, debugging out loud |
| Domain round | 60 min onsite | Perception, autonomy/planning, or systems depth depending on the role |
| System / autonomy design | 60 min onsite | Designing a real-time pipeline that fails safely under sensor loss |
| Behavioral / ownership | 45-60 min, often with a founder | Field grit, bias to ship, how you act when your code moves a real machine |
What the coding rounds feel like
These lean practical over puzzle. You are more likely to get a problem that looks like something from the codebase than a contrived trie question. Think parsing a stream of sensor messages and reconciling timestamps, maintaining a sliding window over a point cloud, or writing a small ring buffer that a real-time thread reads without allocating. If you’re on a C++ track, they will notice whether you know where the copies happen, when a vector reallocates, and why a lock in the hot path is a problem.
A few examples close to how they’re actually asked:
- “Sensor readings arrive out of order with timestamps. Give me the reading closest in time to any query, and keep it fast as readings keep streaming in.”
- “You’ve got a grid of occupancy cells. Find the largest clear region a machine could back into.” (Flood fill wearing work boots.)
- “This function runs every 10ms on the vehicle and it’s occasionally blowing the budget. Here’s the code. Find why and fix it.”
That last kind of prompt matters more here than at most shops. Debugging under a latency budget, talking through what you’d measure, and reasoning about tail behavior instead of average case is the actual job. If you go quiet and just type, you lose signal you can’t get back.
The perception and ML round
For perception roles the questions track the job listings closely: object detection, semantic segmentation, depth estimation, and sensor fusion across camera and lidar. Bedrock also lists behavior-learning methods on some roles, including vision-language-action models and diffusion policies, so if you claim modern policy learning on your resume, be ready to defend the training setup and the failure modes, not just the paper’s abstract.
Good answers here stay grounded in deployment, not benchmarks. If you say you’d add a segmentation head, they’ll ask how you label a construction site where half the objects (rebar, spoil piles, a random parked pickup) don’t appear in any public dataset. If you talk about fusion, they’ll push on calibration drift and time sync between sensors that don’t share a clock. The strongest candidates talk about the boring parts, data pipelines, edge-case mining, and how a model that scores well offline still gets a machine confused when the sun is low and the dust is up.
System and autonomy design
The design round is where the AV heritage shows. A common framing: sketch the pipeline from raw sensors to a command that moves the excavator, and then walk through what happens when a piece of it breaks. Lidar returns garbage in heavy dust. GPS degrades near a tall structure. A perception module misses a person at the edge of the frame. They want to see you design so that the safe action is the default, that degraded sensing narrows what the machine is allowed to do rather than letting it guess, and that a watchdog can bring the machine to a controlled stop when the stack is unsure.
Latency budgets come up a lot. It’s fair game to be asked to allocate a 100ms end-to-end budget across sensing, perception, planning, and actuation, and to defend where you’d spend the milliseconds. Redundancy questions come up too: what runs on a separate safety channel, what you trust to a learned model versus a deterministic check, and how you keep a stop command working even when the fancy part of the stack has fallen over. Answer these like someone who has shipped, not someone who read about ISO safety levels once.
The behavioral round is a real filter
The founder and hiring-manager conversations dig into field grit on purpose. Autonomous construction means someone flies to a site in the summer heat, stands in the dust, and watches the machine do something wrong in a way that never showed up in sim. They want people who find that energizing rather than beneath them. Expect questions about a time you owned a problem end to end, a time production broke and you were the one holding it, and how you make a call when the data is thin and the machine is waiting.
The tell they’re listening for is whether you treat safety as a constraint you design around or a checkbox you resent. Stories where you slowed down, added a guardrail, or killed your own clever idea because it wasn’t safe enough land better than stories about heroics. This is a company where a bad deploy is measured in bent steel, not a rolled-back web build.
Comp, and how to check it
Bedrock hasn’t published bands, and I’m not going to invent a number for you. What’s fair to say: an AV-pedigree, well-funded SF robotics company at this stage tends to pay competitive senior-market base salaries with a meaningful equity grant, and the equity is where the real upside and the real risk both sit. Early-stage equity is a bet on the valuation growing, not cash in hand, and you should price it as such. Ask the recruiter for the specific base, target equity value, and the strike and current preferred price, then sanity-check the base against current data on levels.fyi and comparable AV and robotics startups. Treat any single glassy number online as noise until you’ve triangulated it.
How to prep in the last week
Get your C++ or Python genuinely warm on practical, streaming, and buffer-style problems rather than grinding hard graph puzzles. Re-read your own resume as an adversary and mark every claim you couldn’t defend for ten minutes, then fix those. Pick one perception or autonomy system you actually understand end to end and be able to draw it, including where it fails. And have two or three field-and-ownership stories ready where the point is judgment under uncertainty, because that is the thing this team is buying. The machine doesn’t care how elegant your abstraction is. It cares whether it stops when it should.
Useful next steps:
