Patrick Buehler, PhD

The median Senior SWE at Waymo earns nearly double what the same role at Tesla pays. Senior SWE pay in the automotive industry varies dramatically, and it often comes down to how a company views engineering: as a cost or an investment. Automotive tech sits at the juncture of both software and hardware, and you can see which side a company leans toward just by looking at its pay practices. Companies that start from a software-first ethos tend to extend that compensation philosophy into hardware roles. Those rooted in traditional manufacturing, on the other hand, show more constrained bands. In many ways, how a company began, software or hardware, still defines how it pays, builds, and competes today. We’re bringing back the Senior SWE box-plots, this time for automotive tech, built from recent offer data points in the US. Some insights from the data: Waymo leads the pack with a median of $420K (p25: $380K, p75: $516K). Its 25th percentile clears the 75th at Tesla, Lucid, GM, Zoox, and Rivian, showing clear, consistent top-of-market positioning for autonomous engineering talent. Zoox ($304K) and GM ($302K) cluster in the low-$300Ks with moderate IQRs, implying more standardized offer bands. Tesla shows a lower center (median $213K, p75 $290K) but a long upper tail (max ~$700K), suggesting a mix of outlier grants by department and some volatility based on the stock and refresh grants. Even within the same niche, compensation isn’t just about market value. It’s about philosophy. Some companies reward stability, others risk. And in automotive, that split often mirrors the tech itself: established automakers vs. autonomous pioneers. You can run these reports and visualize distributions like this directly in our Benchmark Tool. Built for compensation and total rewards teams who want to stay ahead of the market. From setting bands to tracking competitor shifts in real-time, our tools help you make data-driven decisions with confidence: https://lnkd.in/gpqPZzcA What category of companies should we chart next? And which range here surprised you most?

263

13 Comments

Tesla just expanded their Robotaxi geofence in ATX. It's now 243 square miles, 3.7x larger than Waymo's. However, the Robotaxi fleet is only 25% of Waymo's size. And there remains human supervisors in the car during rides. I think this is a regulatory thing. Does Tesla's lack of mapping and simpler sensor suite mean that it can move faster and expand more than Waymo? Elon certainly thinks so.... "Tesla autonomous driving might spread faster than any technology ever. The hardware foundations have been laid for such a long time that a software update enables self-driving for millions of pre-existing cars in a short period of time."

230

83 Comments

The autonomous vehicle industry is at a crossroads. Waymo's co-CEO Tekedra Mawakana recently warned that robotaxi companies must go beyond marketing and deliver concrete safety proofs to regulators and the public. With recent incidents involving competitors like Cruise, the pressure is on to establish standardized safety metrics and transparent validation processes. Key challenges include: 1. Data transparency: Companies must share safety test results and incident reports openly. 2. Human-in-the-loop verification: Autonomous systems need rigorous human oversight during edge-case scenarios. 3. Regulatory alignment: Collaborating with agencies to create measurable safety benchmarks. For tech founders building mobility solutions, here's what to prioritize: - Invest in real-world safety testing with third-party validation. - Develop public dashboards to track safety KPIs in real time. - Advocate for industry-wide safety protocols that regulators can audit. The path to mass adoption isn't just about innovation—it's about trust. How will your team balance speed with safety in autonomous tech development? Let's discuss.

9

Diffusion policies forget the past. This method teaches robots to remember… with 3x better performance ⬇️ Robots need memory to act reliably over time, but adding history to diffusion-based policies usually makes performance worse and training more expensive. This new work introduces PTP, an auxiliary loss that brings back context understanding and makes long-history learning fast and effective. Why it matters ✅ PTP helps diffusion policies learn meaningful past-future connections ✅ Enables efficient training with cached short-context embeddings ✅ Boosts performance by 3x while cutting compute costs by over 10x ✅ Adds a test-time scaling trick that checks if the robot is paying attention to its own history Learning long-context robot policies just got way more practical. If we want smarter, more capable robots, this is a big step in the right direction. Thank you so much for sharing this, Marcel Torne Villasevil 🙏 Check out the paper and real-world tasks solved with PTP: Paper: arxiv.org/abs/2505.09561 Website: long-context-dp.github.io Code: https://lnkd.in/dPK2uG9x

99

2 Comments

Over 55M miles of fully autonomous driving data shows the Waymo Driver is making roads safer for everyone, especially for those most at risk – pedestrians, cyclists, and other vulnerable road users. Strong positive impact already and growing exponentially. https://lnkd.in/gQRCCgrN

2,522

62 Comments

🧪🧪EfficientSAM3: Progressive Hierarchical Distillation🧪🧪 👉Bristol announces EfficientSAM3, a family of efficient models built on Progressive Hierarchical Distillation (PHD) that transfers capability from SAM3 to lightweight students. Code coming (in sync with SAM3 release)💙 𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬: ✅3-stage: encoder ⇒ memory ⇒ end-to-end ✅Model zoo of nine EfficientSAM3-PHD students  ✅Flexible accuracy-latency trade-offs for deployment ✅Compact Perceiver-based module trained on SA-V #artificialintelligence #AI #deeplearning #AIwithPapers #LLM 👉Discussion https://lnkd.in/dMgakzWm 👉Paper https://lnkd.in/dnnVUBwC 👉Project https://lnkd.in/d66p5uNN 👉Repo https://lnkd.in/d52YzwtM

359

10 Comments

How distillation will save Waymo and why HD maps and lidar are not a problem Waymo famously has HD 3D maps of areas it drives. Many people think Waymo will fail because 3D mapping the world doesn't scale (it's very expensive to collect and maintain HD 3D maps). Here is what they're missing: Imagine an expensive, huge neural network that is very good at driving, and that requires lots of data as input (e.g. 3D maps + lidar). You deploy it and still potentially lose money because the input data (up-to-date 3D maps) is so expensive. Luckily, there is distillation. A common pattern in distillation is to have a teacher that produces the most accurate predictions you can get, regardless of the cost. You can then use these predictions to train a student model that is lighter and takes less input data. Realistically, you can have a huge teacher that uses 3D maps/lidar/radar/RGB/IMU inputs, and train a lighter student that uses only RGB/IMU/radar. I would bet that in the future even with low-res RGB images we will achieve super-human driving capabilities, using a strong teacher. Note that this means you only need to collect HD 3D maps of few areas to train the teacher + student, and then the student will drive anywhere without 3D data. It won't be necessary to actually map the whole world as some claim. This concept is known as "privileged teacher distillation" or "Learning using privileged information" (LUPI), link to paper in comments. It is very common in teacher/student paradigms. An example is in DINOv3, where the Gram teacher uses 2x larger images than the student. Also, this is not specifically to Waymo, but to any self-driving car company (or any tech in general) that uses temporary crutches. With standard old-school software it would be a nightmare to get rid of such crutches later on once the software is deployed, but deep learning makes it easy

368

13 Comments

“How many takes did that demo require?” That single question exposes the gap between robotics content and robotics reality. Oliver Hsu, from Andreessen Horowitz, nailed the core issue in "The Physical AI Deployment Gap": research is sprinting ahead with VLA models and sim-to-real, while most factory robots still live in a world of fixed routines, tight constraints, and zero tolerance for surprises. The real gap is not intelligence. It’s deployment. Factories don’t care about a perfect 30-second demo. They care about: - running 24/7 - handling variance, drift, and imperfect parts - hitting cycle time - staying stable for months, not minutes At Inbolt, we build for that reality. Not for demos. We are live in production today on hundreds of robots across top manufacturers. Real lines. Real uptime constraints. Our focus is simple: turn CAD-level intent into live robot execution, continuously, fast enough for the robot to adapt while the line keeps moving. If your system needs perfect conditions, it’s not Physical AI. It’s theater. #Robotics #Automation #AI #Manufacturing #Inbolt

37

1 Comment

Tiny neural nets can be very useful in CPU-based rendering. Our latest paper shows how NNs outperform LUTs in accuracy, speed and memory use, and that they can be performant even in batch-size-1 situations. https://lnkd.in/gB9GQ4wG

94

1 Comment

After a few months riding with Waymo, I guess I can officially say I am glad to try the Tesla Robotaxi And yes, I agree on #Optimus. The challenge is fundamentally different. Autonomous driving is hard, but it’s contained: The environment is structured, objective is clear, and the system operates within known rules and constraints That makes it possible to train, simulate, and continuously improve at scale. Humanoid robots are another class entirely. Optimus operates in an open, human-designed world: Unstructured environment, multi-step tasks with shifting goals, and constant interaction with people, objects, and social norms. As work from CSAIL MIT and DeepMind has shown, embodiment dramatically increases complexity: learning transfers poorly, simulation breaks faster, and mistakes carry real physical consequences for humans… A humanoid robot needs a bit more of judgment to be fully capable. Robotaxis prove autonomy can work. Optimus will decide whether autonomy can be trusted. “AI Models are rented. Capability is built.” (…and trust is earned.) HITEC Angeles Investors The HITEC Foundation Unitree Robotics Boston Dynamics

12

3 Comments

🚗 Speed vs. Precision: The Metrics Driving Autonomous Vehicle (AV) Training In the world of Autonomous Driving, the "brain" of the vehicle is only as good as the data it’s fed. As a lead for a human annotation team, I’ve seen firsthand that scaling AV software isn't just about the volume of data—it's about the velocity of high-quality labels. To keep our ML models on track, we focus on three core productivity metrics that balance output with the life-saving precision required for the road: Throughput (Frames/Hour): The baseline of our operation. Whether it’s 3D Cuboids, Semantic Segmentation, or Lane Detection, we track the speed of labeling to forecast model readiness. Quality Consistency (Gold Standard Alignment): Speed is dangerous without accuracy. We measure the "Intersection over Union" ($IoU$) between our annotators and expert "Gold Sets" to ensure pixel-perfect ground truth. Iterative Latency (The Feedback Loop): How fast can we re-annotate "edge cases" (like a pedestrian in a costume or a rare weather event) and get them back to the engineering team? The Human Element: Behind every self-driving car is a dedicated team of annotators making thousands of micro-decisions every hour. Productivity isn't just a number; it’s about building a workflow that empowers people to identify the nuances that sensors might miss. What metrics are you prioritizing in your labeling pipelines? #AutonomousDriving #MachineLearning #DataAnnotation #AI #HumanInTheLoop

21

2 Comments

Are you team Tesla or Waymo? Two very different philosophies in autonomous driving. ✅ Tesla bets on simplicity — a vision-only approach, trusting end-to-end neural networks to interpret the world. ✅ Waymo, on the other hand, builds redundancy into every step — combining LiDAR, radar, cameras, and on-board compute to achieve reliability through sensor fusion. At SEMICON West, Waymo presented a deep look into their system architecture and mission: “Be the world’s most trusted driver.” Key engineering takeaways: - Complete in-house designed 3D sensor suite, integrating: ▪ LiDAR – long-range 360° and short-range perimeter units ▪ Cameras – full-coverage vision system for object detection and classification ▪ Radar – complementing LiDAR under adverse visibility - Central compute node processing hundreds of millions of sensor readings per second - Designed as a safety-critical system with near-zero fault tolerance Tesla or Waymo? Both have one goal: safe autonomy. 🎥 See the short clip below for a look inside Waymo’s sensor suite and I love all the unexpected situations which Waymo is able to handle successfully!

7

7 Comments

How do you turn a single monocular video into a temporally coherent 4D dynamic mesh in seconds? Westlake University, Huazhong University of Science and Technology and Hillbot present “Motion 3-to-4: 3D Motion Reconstruction for 4D Synthesis”. This work tackles the difficult problem of 4D synthesis (3D geometry + motion over time) from monocular input, where depth/motion are ambiguous and high-quality training data is scarce. Motion 3-to-4 proposes a feed-forward approach that decomposes the problem into two parts: static 3D shape generation and motion reconstruction. Using an optional canonical 3D reference mesh, the method learns a compact motion representation and predicts per-frame vertex trajectories to recover full, temporally consistent dynamic geometry. A scalable frame-wise transformer makes the approach robust to varying sequence lengths, enabling efficient 4D generation without slow per-scene optimization. Checkout the links in the comments for more info on the project and the team behind it! #computervision #machinelearning #3dreconstruction #4d #dynamicreconstruction #generativeai #graphics #deeplearning #neuralrendering

98

3 Comments

Last week I wrote about a hacked Toyota Corolla running open-source self-driving software from a $1,000 device by comma.ai In the same days, NVIDIA responded from the #CES2026 stage. They announced Alpamayo: open-source AI models, simulators and datasets for autonomous driving, built around reasoning, not just perception. Chain-of-thought, vision-language-action models that do not only decide what to do, but can explain why. That is a big shift. Autonomy has been optimized for "normal days", but the real problems are rare, messy, human situations that break rigid pipelines. Alpamayo is designed to reason through those. A couple takeaways: 1) Open source is no longer just for geeks. It is becoming a safety feature. If you can see how the system reasons, you can test it, audit it and improve it. That matters much more than flashy demos when regulators, car makers and insurance companies have to trust the technology. 2) Big AI models will train smaller ones that actually drive the car. The smartest models will live in the lab and in simulation. Then their “knowledge” will be copied into lighter versions that can run on real vehicles. Long term leadership in autonomy will likely favor companies that expose their decision logic, not just their performance metrics. Nvidia's press release: https://lnkd.in/dPeUuTWh

25

2 Comments