Robotaxi
How it works
A robotaxi combines an onboard layer and an operations layer. Onboard, a sensor suite (cameras, LiDAR and radar; Tesla uses cameras only) builds a 360° model of the surroundings. A perception module detects and classifies objects, a prediction module forecasts other agents motion, a planner computes a safe trajectory, and a controller actuates steering, throttle and braking. This runs on a high-performance onboard computer in a real-time loop. The vehicle operates only within an approved operational design domain (ODD) backed by HD maps. In unusual situations it can request remote assistance, where an operator provides guidance (not real-time remote driving). A cloud layer handles ride requests, fleet dispatch and rebalancing, charging/maintenance, and data collection for further model training.
Problem solved
Addresses cost, availability and safety of urban transport: it removes the human driver — the most expensive and rate-limiting resource of a ride service — enables 24/7 availability, and, through a fatigue-free, consistent AI "driver", aims to reduce crashes caused by human error.
Components
Cameras, LiDAR and radar providing a 360° view of the surroundings (Tesla uses a camera-only variant).
Detects and classifies objects (vehicles, pedestrians, signs, lanes) and builds a scene representation.
Forecasts other agents motion and computes a safe, rule-compliant trajectory.
Converts the planned trajectory into steering, throttle and braking commands.
High-performance unit (GPU/SoC) running perception and planning networks in real time.
Operators provide guidance in unusual situations; they do not drive the vehicle remotely in real time.
Cloud layer for ride ordering, fleet dispatch and rebalancing, charging and maintenance.
Implementation
Rare, unusual road situations remain the hardest and demand enormous amounts of data.
The service works only within a geofenced area with HD maps; expanding to new cities is costly.
Frequent assistance requests reveal a hidden operating cost and scaling bottleneck.
Incidents (e.g. the 2023 Cruise crash) undermine trust and can lead to permit suspensions.
High vehicle and operating costs mean services still run at a loss.
Evolution
Stanford's Stanley wins the desert race — proof that autonomous driving is feasible.
Autonomy extended to urban traffic with other vehicles and rules of the road.
Google starts its self-driving car effort (later Waymo).
The project becomes the standalone company Waymo under Alphabet.
First commercial robotaxi service in Phoenix (with safety drivers).
Waymo offers public rides with no safety driver in Phoenix.
GM Cruise opens a paid driverless service in San Francisco.
CPUC approves 24/7 paid rides in SF; after a pedestrian-dragging incident Cruise halts operations and recalls 950 vehicles.
Tesla reveals the Cybercab at the "We, Robot" event; Baidu Apollo Go reaches millions of rides in Wuhan.
Tesla launches its Robotaxi service in Austin (initially with a safety monitor).
Waymo reports ~500,000 paid rides and over 4 million miles per week across 10+ cities.
Hyperparameters (configurable axes)
Geographic and conditional envelope in which the vehicle may drive autonomously (area, weather, time of day).
Choice of sensor stack: LiDAR + radar + camera fusion vs a camera-only approach.
Number of vehicles in service; drives availability, wait times and economics.
Number of remote operators per vehicle — a hidden operating cost and scaling bottleneck.
Frequency of disengagements/interventions — a key indicator of system maturity and safety.
Hardware requirements
Onboard autonomous-driving stacks run on high-performance GPUs/SoCs (e.g. NVIDIA DRIVE) for real-time perception and prediction inference.