Vision-Only Perception and the Autonomy Stack Behind It

Removing lidar and radar is the most consequential engineering choice in the programme, and the one whose evidence is hardest to read from outside.

The autonomy stack is where the robotaxi argument is won or lost, and it is also where public evidence is thinnest. Miles driven are not disaggregated by difficulty. Disengagement figures depend on definitions. A video of a smooth drive says nothing about the tail of rare events that decides whether a service can operate without a safety driver.

The six claims here describe sensing, compute, the policy that acts on it, the map it operates within and the loop that improves it. They are inseparable: a perception limit becomes a planning limit, and a planning limit becomes a service-area boundary.

How to read these claims

A robotaxi programme mixes four kinds of statement: a vehicle design shown at an event, a capability demonstrated in a supervised service area, a company target, and a service a member of the public can actually book today. Tesla's pages change as the fleet, software, service areas and terms change, so every claim below is bounded by what was documented on the access date rather than by what may be true in another city or a later release.

Vision-only perception

The source says Robotaxi relies on perimeter cameras without LiDAR or radar.

Cameras provide rich semantic information while making calibration, lighting, occlusion, lens condition and uncertainty estimation central to safety. Sensor philosophy does not substitute for measured performance.

What remains unpublished. Camera count and placement by fleet revision, redundancy, cleaning, low-light performance, any auxiliary sensing and diagnostic thresholds need current documentation.

A fair test. Test glare, darkness, rain, fog, dirty lenses, glass, emergency vehicles and partial camera loss against ground truth and predefined fallback behavior.

Onboard ai compute

The source says Tesla's AI4 or HW4 computer performs real-time neural-network inference in the vehicle.

Onboard compute must sustain perception, prediction, planning, logging and vehicle control within power and thermal limits while retaining a safe response when a processor or input fails.

What remains unpublished. Compute generation, redundancy, memory, power, thermal derating, fault coverage, upgrade path and remote-compute boundary remain incomplete.

A fair test. Profile end-to-end latency and thermal load, then inject processor, camera, time-sync and communication faults on the production configuration.

End-to-end neural networks

The source says neural networks turn raw camera pixels directly into steering, acceleration and braking rather than relying on hard-coded rules.

Learned planning can capture complex road behavior, yet actuation, stability control and fault management still require defined interfaces. Safety claims need behavior and evidence, not a model label.

What remains unpublished. Network boundaries, training mix, action representation, deterministic constraints, validation coverage, rollback and fleet release gates remain proprietary.

A fair test. Document interfaces and evaluate unseen conditions, rare conflicts, forbidden actions, regressions and independent safety monitors across software releases.

3D occupancy understanding

The pasted feature says the system builds a real-time voxel representation of static and moving obstacles.

Occupancy estimates free, occupied and uncertain space without requiring every object to have a name. Planning still needs motion, intent, traffic rules and calibrated uncertainty.

What remains unpublished. Representation, update rate, range, dynamic treatment, uncertainty calibration, hardware differences and current deployment status are unpublished.

A fair test. Compare inferred occupancy with synchronized ground truth for thin, overhanging, partially hidden and moving objects, including false-free-space failures.

Maps and service boundaries

The source says Tesla navigates dynamically without centimeter-accurate HD maps or geographic fences.

Routing, localization, perception priors and service eligibility are separate layers. A dispatch geofence can enforce commercial and regulatory boundaries even if the driving policy reads the road visually.

What remains unpublished. Map data, localization aids, road qualification, boundary enforcement, temporary closures and expansion criteria are not fully described.

A fair test. Attempt routes near service boundaries, changed lanes, construction and stale map data; verify dispatch rejection and safe in-trip boundary handling.

Fleet learning and shadow mode

The source says millions of consumer Teslas continually supply edge cases that update Robotaxi neural-network weights.

Large fleet data can discover rare scenarios but is biased by geography, hardware, customer consent and what triggers upload. More miles do not automatically provide more relevant labels.

What remains unpublished. Robotaxi-specific datasets, trigger logic, consent, retention, regional storage, training cadence, validation gates and regression rates are not public.

A fair test. Audit scenario coverage and lineage from trigger to release, then use held-out routes and staged fleets to measure regressions before broad deployment.

Bottom line

Camera-only autonomy is a defensible engineering position that remains unsettled in public evidence. The right posture is neither the assumption that lidar is mandatory nor the assumption that scale alone resolves the tail.

The disclosure that would move this forward is specific: miles per critical intervention inside a named service area, with weather and time of day broken out, published on a schedule rather than announced after a good quarter.