# Xiaomi Auto World Model 2.0

> Source: https://aiwiki.ai/wiki/xiaomi_auto_world_model_2_0
> Updated: 2026-11-06
> Categories: AI Research, Autonomous Vehicles, Robotics, World Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Xiaomi Auto World Model 2.0." aiwiki.ai, 6 Nov 2026. https://aiwiki.ai/wiki/xiaomi_auto_world_model_2_0
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Xiaomi Auto World Model 2.0** is a streaming [world model](https://aiwiki.ai/wiki/world_model) developed by the [Xiaomi](https://aiwiki.ai/wiki/xiaomi_ai) Auto World Model Team for video generation, autonomous-driving simulation and embodied robotics. Its technical report was submitted to arXiv on November 1, 2026. The system predicts video conditioned on observations, camera poses and controls, and can supply representations to a separate action-generating policy.[1][2]

The report describes generation with one to seven camera views and shows a seven-view driving sequence lasting 45 seconds on internal data. Its driving-policy results are simulation measurements, rather than evidence of a corresponding reduction in road accidents.[1]

## Architecture and inputs

The model starts from the architecture, weights and [variational autoencoder](https://aiwiki.ai/wiki/variational_autoencoder) of Wan2.2-5B. Xiaomi modifies its [diffusion transformer](https://aiwiki.ai/wiki/diffusion_transformer) for multi-view processing and streaming generation. Inputs include a text description, an initial image, camera poses and time-varying world-control signals.[1]

| Component | Function in the reported system |
| --- | --- |
| View-Temporal Attention (VTA) | Alternates processing across time within a view and across camera views at a time step |
| Multi-view rotary position encoding | Distinguishes spatial positions across cameras while retaining a common temporal reference |
| Fused projective rotary position embeddings | Incorporates camera geometry into attention |
| Gated Linear Attention (GLA) | Compresses temporal history into a recurrent state in selected temporal layers |
| World-control injection | Encodes control images with the VAE and concatenates their latents with video latents |
| WorldRec 2.0 | Reconstructs a persistent 3D scene whose renderings condition future video |

The table summarizes the report's architecture.[1] Temporal GLA layers use a state whose size does not increase with the generated sequence length. Periodic softmax-attention layers remain, so the cache property of an individual GLA layer should not be interpreted as eliminating every cache or memory cost in the system.[1]

Control signals have several forms. Driving layouts encode road and traffic information; robot layouts are rendered from a robot model driven by an action sequence. Reconstructed scene renderings provide another control signal for keeping the generated road geometry and appearance consistent as the simulated vehicle changes position.[1]

In this architecture, causal attention restricts a generated chunk to earlier chunks. The paper also uses "causal" to describe learning physical interactions and action consequences. These design choices and the reported evaluations do not establish that every generated event follows real physical laws.[1]

## WorldRec 2.0 scene memory

WorldRec 2.0 constructs static and dynamic 3D Gaussians from observations collected across cameras and time in a single forward pass. Static content remains in world coordinates, while moving objects are represented in their own canonical coordinate systems and transformed to the requested time and pose for rendering.[1]

The reconstruction network uses a small set of learned registers to exchange information among many point queries. Static registers collect background information; dynamic-instance registers connect observations of the same moving object. The resulting scene can be rendered from a new camera pose and fed back to the video generator as a geometric reference.[1]

This addresses a specific problem in long rollouts: video generated without a persistent scene reference can gradually change lane markings, object appearance or cross-camera geometry. The reconstructed content anchors observed regions, while the generator supplies appearance and motion details beyond the observations. The reconstruction is therefore a conditioning source, rather than a complete record of all future scene states.[1]

## Data and training

The team reports approximately 12.6 million training clips. Sources combine general video, driving footage, robotics data, game-engine simulation and generated physical-interaction examples. The data pipeline filters image quality and motion, divides videos into short clips, and uses Qwen3.8 for structured scene descriptions and narrative captions.[1]

| Stage | Training focus |
| --- | --- |
| Pre-training | Single-view text-to-video and image-to-video tasks using broad video and physical-interaction data |
| Mid-training | A progressive mixture of single-view and multi-view tasks with camera-pose and world-control conditioning |
| Post-training | Application-specific driving and robot subsets, followed by consistency distillation and streaming distribution-matching distillation |

These stages retain a common architecture. The report describes separate bidirectional and autoregressive models, and scenario-specific post-trained weights such as seven-view driving and three-view embodied generation. It distills a student that uses two denoising steps per generated chunk.[1]

The corpus has identifiable coverage limits. The report's caption-derived distribution analysis labels 88.8% of physical driving samples as outdoor urban scenes and 72.8% as domestic. Robotics footage comes predominantly from a limited set of indoor environments. These figures describe the sampled corpus, rather than the distribution of situations a deployed robot or vehicle would encounter.[1]

## Streaming inference

For its seven-view inference setup, the team parallelizes generation over eight GPUs: seven handle the camera views, and one handles VAE encoding of world controls. Other optimizations include accelerated rendering, compiled VAE operations, CUDA Graphs, context parallelism and a fixed-size local-attention cache.[1]

An attention-design comparison reports a DiT chunk latency of 6.481 seconds with full softmax attention, 2.299 seconds with VTA and 1.742 seconds with the GLA-based design. The GLA version's FID is 38.08 versus VTA's 37.24, so the efficiency gain comes with a small deterioration in that particular image-distribution metric. These are measurements of the reported setup, not general latency guarantees for arbitrary hardware or camera configurations.[1]

## Video and reconstruction results

Xiaomi evaluates generation, scene perception, camera control and reconstruction separately. Selected results from the team's report are:[1]

| Evaluation | Reported result | Scope |
| --- | --- | --- |
| Internal driving demonstration | Seven views over 45 seconds | Qualitative long-video example |
| nuScenes long-video generation | Single-view FID 25.57 and FVD 125.74 over 211 frames | Lower is better; six-view results were reported separately |
| TriWorldBench generation | TWB-Score 64.79; trajectory accuracy 47.25 | 500 episodes and 50 manipulation tasks |
| Camera-pose control | 78.22 average score | Average of WorldScore and a combined TartanAir/RealCam-Vid set |
| WorldRec 2.0 on Waymo | Full-image PSNR 29.88 dB versus 28.48 for WorldRec 1.0 | Novel-view reconstruction under the report's common protocol |

The TriWorldBench result illustrates why a single overall ranking would be misleading. Xiaomi's model has the highest trajectory accuracy among the methods in the report's comparison, while its aggregate TWB-Score is below CausalWM's 66.04 and WoVR Plus's 65.39. The external numbers use a September 30, 2026 leaderboard snapshot.[1]

## Driving-policy training and planning

The team uses the world model as a simulator for [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) post-training of an end-to-end driving policy. A policy and value head are trained from simulated interaction using PPO. The simulator combines generated observations with trajectory tracking, vehicle dynamics and rewards for factors such as progress, collisions and drivable-area compliance.[1]

For testing, the researchers select 1,000 representative scenes from NAVSIM-derived test data. The report compares the following policies:[1]

| Policy | Collision rate | Scenario completion rate | Minimum time-to-collision | Ego progress |
| --- | ---: | ---: | ---: | ---: |
| DrivoR-E2E | 1.4% | 80.3% | 0.504 s | 81.0% |
| ST-JEPA-E2E | 1.2% | 95.2% | 0.614 s | 84.7% |
| E2E-WM-RL | 0.4% | 96.3% | 0.740 s | 84.5% |

The E2E-WM-RL policy improves the reported collision and completion metrics over ST-JEPA-E2E, while ego progress is slightly lower. These results concern the constructed closed-loop test scenes. They do not measure population-level driving safety or establish performance on all real roads.[1]

A separate [world action model](https://aiwiki.ai/wiki/world_action_model) planning experiment conditions an action model on the world model's latent representations. In a controlled NAVSIM v1.1 navtest comparison using the same action model, data split and training budget, adding the world-model branch raises Predictive Driver Model Score from 94.9 to 95.3. This planning experiment is distinct from the 1,000-scene RL evaluation.[1]

## Robot-policy transfer and evidence limits

For robotic control, the team adds an action-expert diffusion transformer to the video backbone. It reports a 75.0% weighted average success rate on LIBERO-Plus, compared with 49.9% for Fast-WAM and 62.2% for Fast-WAM-Joint in its table. GAM has a higher aggregate score, 85.5%, with a different embodied-pretraining setting.[1]

LIBERO-Plus tests changes in object layout, camera viewpoint, robot starting state, language, lighting, backgrounds and sensor conditions. Its original study predates Xiaomi's model and provides the evaluation context, not independent confirmation of Xiaomi's result.[3]

The November technical report is a preprint and reports the team's own experiments. The project page provides research descriptions and demonstrations; the report's CC BY 4.0 publication license should not be treated as a license for model weights or training data. Neither the demonstrated video horizon nor the simulation scores establish a generally available autonomous-driving product.[1][2]

## References

1. Zhu, Zhenxin, et al. [Xiaomi Auto World Model 2.0: Causal World Models for Real-World Driving and Embodied Intelligence](https://arxiv.org/abs/2611.01677v1). arXiv:2611.01677v1, submitted November 1, 2026. [Full technical report](https://arxiv.org/html/2611.01677v1), especially Sections 2-5 and Tables 3, 8-14.
2. Xiaomi Auto World Model Team. [Xiaomi Auto World Model 2.0](https://xiaomiautoworldmodel.github.io/CausalWorldModel/). Official project page, November 2026. Reviewed November 7, 2026.
3. Fei, Senyu, et al. [LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models](https://arxiv.org/abs/2510.13626). arXiv:2510.13626, first submitted October 15, 2025; revised December 26, 2025.