WeRide is one of China's leading autonomous driving companies. After listing on Nasdaq in 2024, its global hiring tempo accelerated, with Silicon Valley, Guangzhou, and Singapore offices simultaneously hiring SDE / MLE / Robotics Research / Hardware. 1point3acres data shows WeRide OA pass rate is high (~30%), but the real test is the onsite VO: SDE Loop and MLE Loop have very different question types and scoring axes, and many candidates apply to the wrong track and fail outright.
This article focuses on the post-OA Onsite VO flow (for OA topics see our WeRide OA SDE writeup), with a focus on SDE vs MLE differences, perception/planning system design, and a Waymo / Cruise / Pony.ai comparison.
WeRide VO Loop Overview
| Stage | SDE Loop | MLE / Research Loop |
|---|---|---|
| Phone Screen | 60 min coding (LC Medium) | 60 min coding + 30 min ML basics |
| Onsite #1 | Coding (LC Medium-Hard) | Coding (numpy / tensor ops) |
| Onsite #2 | System Design (autonomy stack subsystem) | ML System Design (perception / planning model) |
| Onsite #3 | Project deep-dive | Paper / project deep-dive (must bring 1 representative work) |
| Onsite #4 | Behavioral + Hiring Manager | Behavioral + Tech Lead |
| Onsite #5 | (no SDE round) | Research Discussion / reverse Q |
Key differences:
- SDE: 4 rounds + HM, leaning toward systems / autonomy stack engineering
- MLE / Research: 5 rounds + a mandatory Paper Defense — candidates submit one paper / project upfront and get drilled on it the entire loop
- WeRide doesn't require a PhD — MS + industry experience can apply to MLE, but MS without paper / open-source rarely clears the phone screen
SDE Loop Real Question: System Design
Question — Sensor Fusion + Localization Pipeline
Prompt: an autonomous vehicle has 3 LiDAR + 6 Camera + 1 GNSS/IMU. Design a pipeline that fuses all sensor data to produce the vehicle's 6-DOF pose (position + orientation) at 10ms intervals.
Preferred WeRide Answer Structure
1. Workload
- LiDAR: 3 × 100k points/scan @ 10 Hz
- Camera: 6 × 1080p @ 30 Hz, ~50 MB/s/camera
- IMU: 200 Hz, angular velocity + linear acceleration
- GNSS: 10 Hz, RTK cm-level
2. Time sync layer
- Hardware timestamp (PPS) + PTP / gPTP network sync
- Software buffer + linear interp
3. Preprocessing layer
- LiDAR: motion compensation (IMU-based)
- Camera: undistortion + synchronized capture (HW trigger)
- GNSS: Kalman filter denoise
4. Fusion layer (core)
- Local pose: LiDAR-Inertial Odometry (LIO, e.g., LIO-SAM)
- Global pose: GPS + map matching
- Fusion: Error-State Kalman Filter (ESKF) or factor graph (GTSAM)
5. Output layer
- DDS / ROS2 publish @ 100 Hz
- Failure detection: LiDAR fault → fall back to visual-inertial
Scoring
| Dimension | Penalty | Bonus |
|---|---|---|
| Time sync | Not mentioned | PTP + HW timestamp + < 1ms inter-sensor drift |
| Math | "Use Kalman" | Write out ESKF state vector + jacobian + covariance update |
| Failure | Not mentioned | Degradation paths for single sensor faults |
| WeRide culture | Generic | Cite Apollo / Autoware / WeRide stack differences |
| Engineering | Algorithm only | Latency budget table (each layer < 5ms) |
Trap: many candidates draw an Apollo-style modular pipeline — that's a penalty answer. Since 2024, WeRide has gradually moved to a hybrid end-to-end neural perception + classical planner architecture, and the interviewer wants to see your tradeoff judgment between classical and learning-based methods.
MLE Loop Real Question: ML System Design
Question — Lane Detection Model + Deployment Pipeline
Prompt: real-time (≥ 30 FPS) lane detection on a 4-lane highway, output lane geometry + lane type (solid / dashed / double yellow). Constraints: must run on NVIDIA Drive Orin (254 TOPS) with single-frame latency < 25ms.
Preferred WeRide MLE Skeleton
1. Data
- Sensor: front fisheye + main forward camera
- Scale: 100k frames + 10k long-tail (rain / snow / night)
- Annotation: semi-automated polyline labeling
2. Model architecture
- Backbone: ResNet-34 / EfficientNet-B0 (lightweight)
- Head: LSTR / CondLaneNet (query-based)
- Output: lane parameters (Bezier control points) + class
3. Training
- Loss: line distance + classification + temporal consistency
- Aug: random weather + flip + cutmix
- Train: 4 × A100 for 2-3 days
4. Deployment
- INT8 quantization (TensorRT)
- Batch size 1 (real-time)
- Latency budget: 8ms backbone + 6ms head + 11ms post-proc
5. Online monitoring
- Embedding drift detection
- Long-tail case auto-mining
- A/B canary (10% → 50% → 100%)
MLE Round Bonus Tips
- Quantify: don't say "I used TensorRT for speedup", say "FP32 → INT8 dropped 18ms to 7ms"
- Failure cases: proactively share a past deployment failure (e.g., night-time false positives), root-cause and fix
- Tradeoff judgment: when asked "why not a transformer model?" → answer with latency budget + memory + edge constraints
Project / Paper Deep Dive (mandatory for MLE, occasional for SDE)
MLE Must Bring a "Representative Work"
WeRide MLE recruiting assumes you submit one paper or open-source project, and the entire loop drills into it:
- What's the actual contribution? ("I ran a baseline" doesn't count)
- Difference vs prior art? Citation count?
- If you redid it, what would you change?
- Can this method run in WeRide's actual scenarios? Engineering challenges?
Paper choice strategy:
- CVPR / ICCV / NeurIPS first-author > second-author > advisor-attached papers
- Project should ideally be in autonomy / 3D / perception / planning; pure NLP / vision classification papers score lower as "not aligned with team"
- High-star GitHub open-source can substitute for a paper, but you must articulate your specific contribution (not just fork + minor edits)
Behavioral / Reverse Questions
What WeRide Cares About
- First-principle thinking: AV problems lack canned answers; reason from fundamentals
- Engineering rigor: prototype → production mindset
- Collaboration with hardware / safety: MLE can't only know models — must talk to hardware / safety / firmware teams
- Long-term commitment: AV won't ship in 1-2 years; they want to confirm you're not in "short-term checkbox" mode
High-Frequency Questions
| Question | Recommended Angle |
|---|---|
| "Why autonomous driving?" | Don't say "AI is the future" — give a specific technical interest (e.g., 3D perception uncertainty) |
| "Why WeRide vs Waymo / Cruise?" | Cite WeRide's commercialization (Guangzhou robotaxi + Middle East robobus), tech diversification |
| "Tell me a time you pushed back" | AV safety culture values backbone |
| "Tesla FSD vs LiDAR — your view?" | Must-prep; have a nuanced answer |
3 Reverse Question Directions
- "How does your current model perform on long-tail cases? How do you collect them?"
- "How important is internal simulation? What's your sim-vs-on-road testing ratio?"
- "If I joined, what would be the biggest onboarding challenge in the first 3 months?"
WeRide vs Waymo / Cruise / Pony.ai
| Dimension | WeRide | Waymo | Cruise | Pony.ai |
|---|---|---|---|---|
| HQ | Guangzhou + Silicon Valley | Mountain View | SF (suspended ops 2023) | Beijing + Fremont |
| Commercialization | Guangzhou robotaxi + Middle East robobus | Phoenix / SF / LA robotaxi | Suspended | Beijing / Guangzhou robotaxi |
| MLE Onsite difficulty | ★★★★ | ★★★★★ | (suspended) | ★★★★ |
| H1B sponsor | US office yes (limited) | yes | yes | yes |
| Chinese work env | Guangzhou office strong-Chinese | All English | All English | Beijing office strong-Chinese |
| MLE comp (US NG) | $145-180K + bonus | $180-220K + RSU | (n/a) | $150-180K + bonus |
Recommendation:
- US-China dual presence + visible commercial progress → WeRide
- Most stable / most resources / highest ML talent density → Waymo
- Long-term moonshot / AGI-style vision → Tesla AI / general-robotics companies
FAQ
Q1: Is WeRide SDE or MLE easier to land?
MLE / Research is slightly harder — the bar requires paper / open-source experience, and MS candidates need 1-2 relevant works. SDE is more standardized — LC Medium-Hard coding + system design + projects, leaning on industry experience. No ML experience but strong LeetCode → SDE first.
Q2: Does WeRide MLE require a PhD?
Not strictly. MS + 1-2 CV / Robotics top-conference papers works. MS without papers / open-source rarely passes phone screen. Easiest path for a US candidate: publish CVPR Workshop / ICRA / IROS during school.
Q3: Difference between WeRide US and China offices?
Yes. Silicon Valley (Sunnyvale) office is English-leaning with US benefits; Guangzhou is fully Chinese, faster pace, more OT. US OPT/H1B candidates default to Silicon Valley, but the hiring committee evaluates "willingness to rotate to Guangzhou for 1-2 years" during packet review — unwilling candidates are down-ranked.
Q4: How fast does WeRide give VO results?
After the full onsite, 3-7 days for a recruiter call with verbal feedback, offer letter within 1-2 weeks. If 2+ weeks pass with no signal, 80% chance you're on hold (hiring committee couldn't reach consensus, or offer budget is queued) — proactively follow up the recruiter.
Q5: Does WeRide sponsor H1B?
US office: yes, with annual limits. WeRide prioritizes PhD / Senior roles in main H1B rounds; NG SDE / MLE typically bridge with OPT + STEM extension and apply for H1B later. No H1B transfer to a new office — you must onboard first before WeRide files H1B.
Q6: I have offers from both WeRide and Waymo — how do I choose?
Generic guidance:
- Waymo suits: top ML resources, stable big-tech, English-only, long-term R&D
- WeRide suits: close-up view of commercialization (robotaxi already monetized), US-China flexibility, faster early-career promotion
Comp: Waymo is 15-25% higher (especially RSU), but WeRide post-IPO RSUs have better liquidity (Waymo is an Alphabet subsidiary — no separate stock).
Contact
If you're prepping WeRide, Pony.ai, Waymo, Cruise, Tesla Autopilot — autonomy SDE / MLE / Research — OA is not the variable; onsite system design and paper deep-dives are where 60% of candidates fall short. We've curated WeRide 2025-2026 Onsite real questions + autonomy system design templates + paper defense prep checklist.
Add WeChat Coding0201, get the WeRide / autonomy onsite bank and book mocks.
- Email: [email protected]
- Telegram: @OAVOProxy