# GAIA-4 (Wayve)

> Source: https://aiwiki.ai/wiki/gaia_4_wayve
> Updated: 2026-08-04
> Fact-checked: 2026-08-04
> Categories: AI Models, Autonomous Vehicles, Computer Vision, Generative AI, World Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "GAIA-4 (Wayve)." aiwiki.ai, 4 Aug 2026. https://aiwiki.ai/wiki/gaia_4_wayve
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**GAIA-4** is a multimodal generative [world model](https://aiwiki.ai/wiki/world_model) for closed-loop autonomous driving simulation, announced by the British self-driving company [Wayve](https://aiwiki.ai/wiki/wayve) on 3 August 2026 as the latest generation of its GAIA family.[1][2] Where Wayve positioned the earlier [GAIA-2](https://aiwiki.ai/wiki/gaia_2_wayve) and [GAIA-3](https://aiwiki.ai/wiki/gaia_3_wayve) models primarily as controllable generators of synthetic driving video and as engines for offline evaluation, GAIA-4 puts the company's end-to-end driving model, the Wayve AI Driver, directly inside the simulation loop: the driving model's decisions change the sensor inputs it receives next, so a recorded scene becomes the starting point for new simulated futures rather than a fixed replay.[1] Wayve describes GAIA-4 as the world model at the core of "Simulation 2.0", its proposed framework for evaluating an end-to-end driving system as a single behavioural unit, and says GAIA-4 brings generated radar into the simulation alongside camera video, which the company calls "an industry-first for autonomous driving simulators".[1]

The announcement was made through a research post on Wayve's blog titled "GAIA-4: Multimodal World Models Powering Closed-Loop Simulation for Safe and Scalable Autonomy" and an accompanying post on X, in which the company summarised the release as "transforming world modeling from scene generation into a way to measure end-to-end safety at scale".[1][2] Unlike the GAIA-1 and GAIA-2 releases, which were accompanied by technical reports, the GAIA-4 launch post does not link a technical report or paper, and it does not state the model's parameter count, architecture, training compute or training data volume.[1] The quantitative claims in the post, including a reported 2.5 times improvement in how faithfully the model preserves a recorded scene, are Wayve's own and had not been independently verified at the time of writing.[1]

## Key facts

| Attribute | Detail |
| --- | --- |
| Developer | [Wayve](https://aiwiki.ai/wiki/wayve), London based autonomous driving company |
| Announced | 3 August 2026, via research blog post and X post |
| Type | Multimodal generative world model for closed-loop driving simulation |
| Role | Core world model of Wayve's "Simulation 2.0" evaluation framework |
| System under test | Wayve AI Driver, the company's end-to-end driving model |
| Modalities | Camera video and radar, generated coherently by the same model |
| Signature capabilities | Closed-loop simulation with the AI Driver in the loop; "world-on-rails" counterfactuals; reactive agents; replay of safety operator interventions |
| Grounding | Scenes generated directly from recorded fleet sensor data, without HD maps, scene graphs, prompts or a separate annotation stack |
| Predecessors | GAIA-1 (2023), [GAIA-2](https://aiwiki.ai/wiki/gaia_2_wayve) (March 2025), [GAIA-3](https://aiwiki.ai/wiki/gaia_3_wayve) (December 2025) |
| Technical report | None linked; parameters and training data not disclosed |

## Why does end-to-end driving need a new simulator?

Wayve frames GAIA-4 as a response to a structural problem created by its own approach to autonomy, which it calls AV2.0: replacing the modular sense-plan-act stacks of earlier autonomous vehicle programmes with a single end-to-end neural network that maps raw sensor inputs directly to driving decisions.[1] In Wayve's argument, end-to-end learning is what allows a driving system to generalise across cities, sensors and vehicle platforms, but the same property, one learned system rather than separable parts, breaks the tools the industry has historically used to prove safety, including simulation.[1]

The blog post distinguishes between two generations of simulation. What Wayve calls Simulation 1.0 was designed for modular systems: when perception and planning are separate components, each can be validated on its own terms, with detection accuracy metrics for perception and trajectory error metrics for planning. End-to-end driving models are not separable in this way. There is no intermediate perception output to score and no planning module to test in isolation; behaviour emerges from a single network, so the only meaningful unit of evaluation is the whole system, judged by how it behaves.[1]

Wayve calls the alternative Simulation 2.0, which it defines as "simulation purpose-built to evaluate a driving model as one behavioral system", and derives two requirements from it. First, the simulator must faithfully reproduce the full sensor experience the driving model consumes, across all modalities, because an end-to-end model drives on raw sensor data. Second, replay and open-loop evaluation are not enough: a fixed recording reveals what a model would predict, but its predictions never change what happens next. To learn what the vehicle would actually do, the simulator has to close the loop and let the model's decisions change what it sees next.[1]

Wayve also connects the argument to the economics of safety cases: as a driving model improves, the number of miles needed to validate it sharply increases while almost all of those miles are uneventful, so real-world testing alone cannot carry an autonomy safety case.[1]

## How does GAIA-4 close the loop?

GAIA-4 turns a recorded driving scene into the starting point for a new simulated run with the Wayve AI Driver back in the loop. As the driving model makes different decisions from those made in the original recording, the sensor inputs it receives change accordingly: if the AI Driver brakes, the simulated viewpoint slows, and if it steers, the scene it sees next shifts with it. Rather than playing back a fixed recording, the simulator generates the futures those decisions create, including futures that never occurred in the real world, while keeping the underlying scene anchored to what actually happened, which provides a deterministic baseline to measure against.[1]

A distinguishing design choice is how the world is constructed. According to Wayve, GAIA does not rely on HD maps, hand-built scene graphs, prompts or a separate annotation stack; it generates the scene directly from recorded sensor data, which lets the simulator hold the world faithful to what actually happened while only the ego vehicle changes.[1]

Wayve presents this as changing the scaling behaviour of validation: GAIA-4 can run many candidate driving models through a library of recorded scenarios in closed loop without re-driving any of them, and because each predicted action feeds back into the generation, every run tests what the model would actually do. In the company's summary, "Validation stops scaling with road miles and starts scaling with compute."[1]

## World-on-rails and reactive agents

The launch post introduces "world-on-rails" as the default constraint for safety validation. Under world-on-rails, the AI Driver's actions drive the simulation forward, changing its position, its viewpoint and the sensor stream it receives, but every other agent in the scene keeps the exact behaviour it showed in the real-world log. Other road users do not react to the ego vehicle: a car that was going to enter an intersection still enters it, and a cyclist crossing the street still crosses.[1]

Wayve argues this constraint gives the simulation two properties that matter for validation. First, other road users stay grounded in real behaviour, because they continue to do what they actually did in the recorded drive. Second, the evaluation stays conservative: no vehicle, pedestrian or cyclist changes its behaviour in response to the AI Driver, which prevents the simulator from softening a safety-critical outcome by having the world accommodate the ego vehicle. Under this constraint, Wayve says it can assess the safety outcome of any scenario, including scenarios in which a vehicle safety operator intervened during the original drive.[1] The post illustrates the point with a junction scene generated with and without the constraint: without world-on-rails, the model can generate a forgiving future in which no car appears in the junction, which completely changes the safety criticality of the event.[1]

World-on-rails is described as a capability that can be switched on rather than a limitation of the model. Relaxing the constraint enables what Wayve calls reactive agents, in which selected road users no longer simply replay the recording but respond to what the ego vehicle does, while the rest of the scene continues to follow the original log. Wayve positions reactive agents as an extension for testing interactions that only real reciprocity can probe, such as negotiated merges and give-and-take at junctions, with world-on-rails remaining the conservative baseline for safety validation.[1]

Wayve reports that training GAIA for the world-on-rails reconstruction task improves how faithfully it preserves the recorded world by 2.5 times, which the company says turns the generator into a system stable enough to be used as a closed-loop simulator. The post does not describe the metric or baseline behind the 2.5 times figure, and the claim has not been independently verified.[1]

## Counterfactuals: interrogating the AI Driver

Closed-loop simulation grounded in real recordings is presented as an introspection and debugging tool as much as a validation tool: take any real moment, change one thing, and watch what the AI Driver would have done, turning an on-road event into a repeatable diagnosis of the model's behaviour. Two questions recur in the post: what would the AI Driver have done if the vehicle safety operator had not intervened, and how does a new model behave in exactly the situation that tripped up the previous one.[1]

The post shows examples in which a safety operator intervened during a real drive, alongside GAIA-4 counterfactuals in which no intervention occurs and the AI Driver resolves the situation itself; Wayve states that in the example shown, the AI Driver's own choices in simulation maintained a wider safety margin than the operator intervention did. These are Wayve-selected examples rather than results from a disclosed evaluation set.[1]

## Multimodal generation: radar in the loop

GAIA-4 extends the GAIA line beyond camera video by generating radar alongside cameras, which Wayve calls "an industry-first for autonomous driving simulators".[1] The company's argument is that a production driving system relies on several sensing modalities, such as cameras, radar and lidar, and that simulation becomes far more valuable when those signals are generated coherently as part of the same world rather than assembled separately.[1]

Radar matters in this framing because it provides redundancy and robustness in adverse conditions such as fog, spray and low light, where camera visibility is poor. In the highway example shown in the post, generated radar tracks a lead vehicle and an oncoming car with colour-coded Doppler measurements showing each object's velocity relative to the ego vehicle, and the detections are produced by the same world model that generates the video rather than being added separately afterwards.[1]

Wayve also presents an experiment intended to show that the generated radar carries decision-relevant information. A highway scene in which a vehicle failed to slow down is run in closed loop first with cameras only and then with synthetic radar included. According to the post, adding radar measurably shifts the closed-loop trajectory and the vehicle decelerates appropriately, which the company offers as evidence that generated radar is a usable model input rather than a visual overlay.[1]

## How does Wayve measure simulator fidelity?

Because the simulator itself is a learned generative model, its trustworthiness needs evidence, and the launch post describes a three-level framework for measuring what Wayve calls behaviour fidelity.[1]

| Fidelity level | Question asked | Method described |
| --- | --- | --- |
| Outcome fidelity | Do simulated runs reproduce the outcomes observed in the real world, above all on which scenarios are safety-critical? | Simulated outcomes are graded with the same metrics Wayve uses to evaluate real-world driving |
| Closed-loop fidelity | Are the system's outputs in simulation consistent with its on-road driving? | The trajectory the AI Driver plans in simulation is compared frame by frame with the trajectory it planned on the road; with the same model the two should agree, visualised as a tight "ribbon" of convergence |
| Component fidelity | Are the individual simulator components valid? | Per-component tests of 3D realism, geometric consistency, vulnerable road users, traffic signs, lane features and radar fidelity |

For component fidelity, Wayve exploits the fact that the original recording captured what was actually happening: every element of the scene can be geometrically projected from the original viewpoint into a new viewpoint, and GAIA's generation can be checked against that projection. The post describes perturbing the ego vehicle's pose and verifying that other vehicles and lane markings remain stable, and measuring how far the ego vehicle can diverge from the original path while traffic signs and roadside objects remain legible and correctly placed. Wayve singles out traffic signs as particularly important for driving and typically hard to recreate in simulation.[1] For radar, one described metric measures whether radar points continue to track key elements of the scene, such as the vehicle in front, across conditions; when the ego vehicle is made to accelerate or decelerate, Wayve says GAIA generates in-distribution radar in response, with realistic changes in Doppler and coverage as the ego vehicle approaches vehicles ahead.[1]

The post presents this framework qualitatively with video examples, without numerical results, benchmark scores or comparisons against other simulators.[1]

## Position in the GAIA lineage

GAIA-4 is the fourth public generation of Wayve's GAIA (Generative AI for Autonomy) world model family. The launch post frames it as the product of eight years of world model research, starting with what Wayve describes as the first world model for autonomous driving, published as "Dreaming about driving" in September 2018.[1][3]

| Generation | Public release | Reported scale | Primary positioning |
| --- | --- | --- | --- |
| GAIA-1 | June 2023 preview; technical report 2023 | 9 billion parameters, trained on about 4,700 hours of London driving data | Proof-of-concept autoregressive generative world model conditioned on video, text and action[4] |
| [GAIA-2](https://aiwiki.ai/wiki/gaia_2_wayve) | 26 March 2025 | Latent diffusion world model with extensive domain-specific conditioning | Controllable multi-camera synthetic scenario generation across UK, US and German driving data[5][6] |
| [GAIA-3](https://aiwiki.ai/wiki/gaia_3_wayve) | 2 December 2025 | 15 billion parameters, pre-trained on ten times more data than GAIA-2 | Offline evaluation and validation, "from simulation to evaluation"[7] |
| GAIA-4 | 3 August 2026 | Not disclosed | Multimodal closed-loop simulation with the AI Driver in the loop, measuring end-to-end safety[1] |

The through-line Wayve draws across the generations is a shift in the consumer of the model's output. GAIA-1 and GAIA-2 were largely judged on the realism and controllability of the video they generated, GAIA-3 reframed generation as an instrument for offline evaluation, and GAIA-4 completes the turn by making the driving model itself the audience, with the AI Driver's own actions steering the generation.[1] Wayve argues that end-to-end learning applies to safety and simulation just as it applies to driving, and that embodied AI safety improves faster with learned representations than with hand-engineered ones.[1]

## Significance and open questions

Wayve presents GAIA-4 as the foundation of a new kind of safety case for end-to-end autonomy. In the company's framing, any team credibly building end-to-end autonomy inherits a system that can no longer be taken apart and tested in pieces, and the simulation tools built for modular stacks cannot validate it. Closed-loop simulation grounded in real recordings is offered as the answer: every intervention can be replayed and interrogated, every model change can be measured against the same real scenarios, and every sensor the vehicle depends on can be put under test, turning the end-to-end AI Driver from a supposed "black box" into a system Wayve says can be examined, validated and trusted.[1] The post ties this to the company's commercial strategy, describing mapless end-to-end AI as the most realistic path to autonomy at geographic scale and closed-loop simulation as what makes it possible to scale safely "across millions of vehicles" with Wayve's automaker partners.[1] GAIA-4 is described as one building block of Simulation 2.0, Wayve's broader framework for evaluating the AI Driver for commercial deployment, with more components to come.[1]

Several aspects of the system remain undisclosed or unverified. The launch post contains no parameter count, architecture description, training data volume, resolution figures or numerical evaluation results, and no technical report or arXiv paper is linked from it, in contrast to the GAIA-1 and GAIA-2 releases.[1] The central claims, including the 2.5 times fidelity improvement, the industry-first status of generated radar and the agreement between simulated and on-road behaviour, are all Wayve's own statements about its own product and had not been independently reproduced or benchmarked by external researchers at the time of the announcement. Lidar, which the post names as one of the modalities a production system relies on, is discussed as motivation but is not claimed as a generated modality in the launch material, which demonstrates cameras and radar.[1]

## ELI5

Imagine a dashcam video of a tricky moment on the road. Older versions of Wayve's GAIA could redraw that video in different weather or add extra cars, like a very smart video editor. GAIA-4 goes further: it hands the steering wheel inside the video to Wayve's robot driver. If the robot brakes or turns differently from what happened in real life, GAIA-4 draws the new version of events those choices would cause, while everyone else on the road keeps doing exactly what they really did. That lets Wayve safely ask questions like "what would the robot have done if the human safety driver had not grabbed the wheel?" on computers instead of real streets. GAIA-4 also imagines radar signals, not just pictures, so the robot gets all the senses it would have in a real car.

## See also

- [Wayve](https://aiwiki.ai/wiki/wayve)
- [GAIA-3 (Wayve)](https://aiwiki.ai/wiki/gaia_3_wayve)
- [GAIA-2 (Wayve)](https://aiwiki.ai/wiki/gaia_2_wayve)
- [World model](https://aiwiki.ai/wiki/world_model)

## References

1. Wayve. "GAIA-4: Multimodal World Models Powering Closed-Loop Simulation for Safe and Scalable Autonomy." Research blog, 3 August 2026. https://wayve.ai/thinking/gaia-4/
2. Wayve (@wayve_ai). "Introducing GAIA-4, the latest evolution in our pioneering line of world models." X post, 3 August 2026. https://x.com/wayve_ai/status/2084282138795397353
3. Wayve. "Dreaming about driving." Research blog, 5 September 2018. https://wayve.ai/thinking/dreaming-about-driving/
4. Wayve. "Scaling GAIA-1: 9-billion parameter generative world model for autonomous driving." Research blog, 2023. https://wayve.ai/thinking/scaling-gaia-1/
5. Wayve. "GAIA-2: Pushing the Boundaries of Video Generative Models for Safer Assisted and Automated Driving." Research blog, 26 March 2025. https://wayve.ai/thinking/gaia-2/
6. Russell, L., Hu, A., Bertoni, L., Fedoseev, G., Shotton, J., Arani, E., and Corrado, G. "GAIA-2: A Controllable Multi-View Generative World Model for Autonomous Driving." arXiv:2503.20523, 26 March 2025. https://arxiv.org/abs/2503.20523
7. Wayve. "Wayve launches GAIA-3, advancing world models from simulation to evaluation." Press release, 2 December 2025. https://wayve.ai/press/wayve-launches-gaia3/