SoftTech AI Registry
Waymo · Waymo Driver

Model-Derived Driving Trajectory Generation

The production Waymo Driver uses learned AI models to generate vehicle trajectories for real-time autonomous driving. Distilled Driver models run onboard and produce trajectories that are checked by a separate safety-validation layer before physical driving execution; ordinary driving does not require individual human approval.

Recorded characteristics

Function
Controlled object: one generated trajectory for one Waymo vehicle. Lane changes, yielding, merging, turning, braking, acceleration and steering manoeuvres are manifestations of this trajectory-generation authority and are not recorded as separate capabilities. Causal chain: sensor/environment inputs → learned world/model representation → learned/model-derived trajectory generation → generated/proposed ego-vehicle trajectory → independent onboard safety validation → downstream vehicle-control system → physical vehicle movement. The exact transformation from the validated trajectory into individual actuator commands is not publicly established; no claim is made that the model commands steering, brake pressure, accelerator or wheel torque. Model authority: generation of the ego vehicle's own trajectory. Waymo states the Foundation Model's encoders feed the World Decoder, "which uses these inputs to predict other road users behaviors, produce high-definition maps, generate trajectories for the vehicle, and signals for trajectory validation"; Teacher Driver models "are trained to generate safe, comfortable, and compliant action sequences" and are distilled into "more efficient Student models, optimized for real-time onboard deployment". Prediction of other road users is not the qualifying authority; perception, object detection, prediction, mapping and localisation are not used as evidence of action authority. Physical consequence: the model-generated trajectory, after onboard validation, guides the autonomous vehicle's subsequent physical movement through the road environment.
Data access
Onboard multimodal sensor inputs (cameras, lidar, radar) and HD maps used as a prior/input. Exact production feature set not publicly established.
Actions
Can take actions
External actions
Yes
Human confirmation
Not required
Permission basis
Not established
Administrative control
Generative ML Driver = trajectory-generation authority. Validation/safety architecture = constraint, checking and veto authority: an "independent onboard validation layer", "a separate, AI-based safety system that monitors every trajectory proposed by the Waymo Driver", checking plans "against hard physics-based constraints and traffic laws"; "If the AI proposes a path that violates a limit or risks a collision, the validation layer acts as a hard backstop." The validator does not ordinarily generate the trajectory. Standing constraints bounding (not determining) the ordinary trajectory: physical feasibility, traffic laws, collision avoidance, mapped/operational geography, road environment, route/destination objectives. Human confirmation not_required applies to ordinary trajectory generation/execution only. Remote Assistance does not remotely drive the vehicle: Waymo states "Waymo's service does not rely on remote drivers"; RA responds to requests "initiated by the Waymo Driver" and provides "advice which the system can decide to use or reject". Humans do participate in Waymo operations (Remote Assistance, Event Response, emergency coordination, post-collision procedures); these are not credited to this capability. Outside action yes: the consequence is movement of the vehicle through the physical road environment. Permission basis not_established: ride requests, Waymo's ownership/control of the vehicle, road permission and regulatory authorisation are not Registry software permission mechanisms. Default state not_established: no customer-deployment toggle maps to the Registry's default-state methodology.
Default state
Not established
Availability
Current fully autonomous (rider-only) Waymo operation; Waymo reports more than 200 million fully autonomous miles (August 2026). Fleet scale is deployment context only, not action blast radius; blast radius is one generated trajectory for one Waymo vehicle. Deployment parity across every geography and vehicle platform is not publicly established.
Licensing
Not applicable: operated by Waymo; no customer licensing of the Driver established.
External model or provider
Waymo's own models: Waymo Foundation Model (encoders feeding the World Decoder), with Driver capabilities distilled from Teacher models into Student models optimised for real-time onboard deployment. VLM trained with Gemini contributes semantic signals; its exact contribution is not publicly established.
Limitations and uncertainty
No automated monitor created because current Waymo website terms restrict automated data extraction/scraping. Learning: no continuous online learning by an operating vehicle is claimed; Waymo describes an offline cycle (operational data and Critic findings → scenario extraction/auto-labelling → retraining → simulation → safety-framework validation → deployment); continuous production learning is not publicly established. Reversibility: future trajectory plans may be superseded; an in-progress manoeuvre may be changed where physically possible; the vehicle may stop or alter course; physical movement already completed cannot be undone. Exclusions: (1) perception alone; (2) object detection; (3) classification; (4) scene understanding alone; (5) prediction of other road users alone; (6) mapping; (7) localisation; (8) passenger route selection; (9) EMMA; (10) S4-Driver; (11) simulation-only models, including the Waymo World Model simulator; (12) research-only planning systems; (13) Critic evaluation/training processes; (14) Remote Assistance advice; (15) Event Response; (16) safety-validator authority; (17) backup collision avoidance; (18) safe-stop/fallback authority; (19) exact actuator control; (20) direct steering-command generation; (21) direct brake-command generation; (22) continuous online learning; (23) separate lane-change/yield/merge/braking capability claims; (24) redundant compute, steering, braking and backup power as safety architecture, not model output. Research systems carry no classification-sensitive field. Not publicly established — Model/planning: (1) exact onboard Student-model architecture; (2) parameter count; (3) exact model family/version; (4) precise World Decoder architecture; (5) exact trajectory-output representation; (6) candidate-trajectory count; (7) candidate-generation architecture; (8) exact scoring/ranking mechanism; (9) exact numerical objective; (10) confidence/uncertainty per trajectory; (11) feature attribution; (12) exact VLM contribution; (13) precise learned/structured representation boundary; (14) exact contribution of conventional engineered planning code. Validation/safety: (15) exact validation algorithm; (16) exact production RL role in validation; (17) exact generative-reasoning mechanism; (18) exact veto thresholds; (19) whether the validator modifies or only rejects a trajectory; (20) complete fallback behaviour after validation failure; (21) complete priority relationship between Driver, validator and backup collision system. Control: (22) trajectory-to-steering transformation; (23) trajectory-to-braking transformation; (24) actuator-command generation; (25) control-loop frequency; (26) actuator-authority hierarchy. Operations: (27) universal communications-loss behaviour; (28) all Remote Assistance request triggers; (29) exact effect of Remote Assistance advice on subsequent planning; (30) complete sensor-failure hierarchy; (31) complete compute-failure hierarchy; (32) universal degraded-mode behaviour; (33) deployment parity across every geography; (34) deployment parity across every vehicle platform. Learning/audit: (35) online learning during production driving; (36) model version associated with each trajectory; (37) per-trajectory model trace; (38) public decision-log schema; (39) log retention; (40) external availability of trajectory logs; (41) complete per-action explanation.

Evidence