---
topic: system-design
author: Crashtech Editorial
published: Aug 17, 2026 · read: 5 min
---

Written Down or Computed Live: Comparing Waymo and Tesla on Sensing, Planning, and Safety

Lidar and prior roadgraphs versus end-to-end vision neural networks: the architectural divergence between Waymo and Tesla.

– –

At 40 miles per hour, a moving automobile travels approximately 60 feet every single second. Inside that one-second window, an autonomous driving system must perceive surrounding objects, predict their future trajectories across multiple possibilities, plan a collision-free path, and actuate steering and braking.

If an unusual scenario appears (e.g., a mattress falling off a truck on a highway, or a construction worker waving traffic into the opposite lane), the software must resolve it without hesitation.

In the autonomous vehicle industry, two technological giants have converged on diametrically opposed software architectures: Waymo and Tesla.

This competition is not merely a debate over lidar versus cameras; it represents a profound philosophical divergence: Should reality be written down in advance, or computed live by neural networks?

+-------------------------------------------------------------------------+
|                  THE CORE ARCHITECTURAL DIVERGENCE                      |
|                                                                         |
|  [Waymo: Structured Modular Robotics]                                   |
|  - Multi-Modal Sensor Fusion (Lidar + Radar + Cameras)                  |
|  - Pre-computed HD Prior Roadgraphs (Centimeter-accurate maps)          |
|  - Explicit Trajectory Prediction & Rule-Based Safety Envelope          |
|                                                                         |
|  [Tesla: End-to-End Neural Network]                                     |
|  - Vision-Only Sensing (8 Optical Cameras)                              |
|  - Zero Prior HD Maps (Navigates via GPS & Live Bird's-Eye View)        |
|  - Monolithic Neural Weight: Pixels In -> Actuation Controls Out        |
+-------------------------------------------------------------------------+

Layer 1: Sensing Mechanics

The first architectural boundary is how the vehicle digitizes physical space.

Waymo: Multi-Modal Sensor Fusion

Waymo equips each vehicle with a suite of complementary sensors:

  • Rotary and Solid-State Lidar: Emits millions of laser pulses per second, directly measuring the time of flight to calculate precise 3D point clouds independent of lighting or shadow conditions.
  • Imaging Radars: Penetrate heavy rain, fog, and dust, measuring Doppler velocity directly.
  • High-Resolution Cameras: Classify semantic labels (traffic lights, temporary signs, emergency vehicle lights).

Because lidar measures physical photons directly, Waymo’s system never has to guess whether a dark rectangular patch on the highway is a shadow or a solid concrete block.

Tesla: Vision-Only (Pure Neural Reconstruction)

In 2021, Tesla eliminated radar and ultrasonic sensors from its vehicles, committing completely to optical cameras:

  • Eight standard RGB cameras surround the car.
  • The Neural Depth Engine: A deep convolutional and transformer backbone infers depth, velocity, and 3D geometry purely from optical video streams.
  • The Philosophy: Humans navigate the world using biological eyes and a neural brain without emitting laser beams; therefore, silicon computers can achieve the same with cameras and synthetic neural networks.

Layer 2: Representation and HD Mapping

How does the car know where the road is?

Waymo: The Authoritative Roadgraph

Waymo maps its operating domains in advance using centimeter-accurate High-Definition (HD) Maps:

  • Every curb, crosswalk, stop line, traffic signal link, and speed limit is pre-surveyed and stored in local memory.
  • When driving, Waymo does not need to deduce which traffic light controls its specific lane; the mapping database pre-links the light to the lane.
  • Real-time sensors only need to detect dynamic objects (pedestrians, cars, cyclists) and identify deviations from the map (construction zones).

Tesla: Live Bird’s-Eye-View (Occupancy Networks)

Tesla rejects HD maps, arguing they do not scale globally and become dangerous when construction alters the physical layout:

  • Tesla vehicles navigate using standard navigation GPS (like Google Maps) combined with live Occupancy Networks.
  • The neural network continuously generates a 3D voxel representation of drivable space in real time.
  • If a road has no lane markings or is covered in snow, the neural net predicts the plausible path dynamically.

Layer 3: Prediction and Planning

Once objects are perceived, how does the vehicle decide where to steer?

Waymo Modular Pipeline:
[Lidar/Camera] -> [Perception Tracker] -> [Multi-Future Predictor] -> [Rule-Based Planner]
                                                                                |
                                                                                v
                                                                   [Hard Safety Invariant Gate]

Tesla End-to-End FSD V12+:
[Raw Video Streams] =============================================> [Steering / Throttle / Brake]
                    (Billions of End-to-End Transformer Weights)

Waymo: Explicit Trajectory Evaluation and Safety Envelopes

Waymo splits prediction and planning:

  1. Multi-Future Prediction: For every detected car and pedestrian, an ML model generates multiple probability-weighted trajectory trees over a 5-second horizon.
  2. Trajectory Optimization: A deterministic planner evaluates candidate paths against safety constraints (safe following distance, acceleration comfort).
  3. Hard Safety Envelope: If an ML model proposes an aggressive lane change, a deterministic safety validator enforces hard physical boundaries (deceleration limits, collision boundaries) and overrides the model if an invariant is violated.

Tesla: End-to-End Neural Policy (FSD V12)

Prior to version 12, Tesla used 300,000 lines of C++ code to implement heuristics and state machines. With FSD V12, Tesla replaced the algorithmic planner with an end-to-end neural network:

  • Video frames feed into an enormous transformer network trained on millions of hours of human driving video clips.
  • The network directly emits control outputs: steering angle, acceleration, and braking.
  • There are no explicit C++ if/else statements for stop signs; the network learned to stop because human drivers in the training data stopped at stop signs.

The Validation Divergence: Real-World Safety Data

How do we evaluate which architecture is safer?

The two companies measure safety using completely different denominators:

  • Waymo (Rider-Only Driverless Operations): Evaluates millions of fully driverless commercial miles (with zero safety drivers) against human crash benchmarks in Phoenix, San Francisco, and Los Angeles. Waymo reports an 85% reduction in injury-causing crashes compared to human drivers in the same geographies.
  • Tesla (Supervised Fleet Scale): Leverages millions of consumer vehicles gathering billions of miles of real-world fleet data daily. However, because a licensed human driver is legally required to supervise the car and intervene during errors, separating the vehicle’s autonomous capability from human intervention remains statistically challenging.

The Systems Tradeoff

DimensionWaymoTesla
Sensing RedundancyExtreme (Lidar + Radar + Vision)Minimal (Cameras only)
Hardware CostHigh ($40,000+ per vehicle sensor pod)Low (~few hundred dollars per car)
Geographic ScalabilityGeofenced (Requires pre-mapped cities)Global (Runs on any road worldwide)
System PredictabilityHigh (Hard deterministic safety gates)Probabilistic (End-to-end neural weights)
Commercial ModelRobotaxi Fleet Operator (Autonomous Ride)Consumer Vehicle Software Add-on

Waymo chose the path of guaranteed safety within bounded geographic envelopes. Tesla chose the path of unconstrained global scalability. The race between them will settle whether physical safety in complex distributed environments can ever be trusted entirely to probabilistic neural weights.

Advertisement

/* Comments */