We present Xiaomi Auto World Model 2.0, a unified streaming causal world model for Physical AI. It models and generates causal physical evolution in the real world, and unifies heterogeneous conditioning inputs for view-consistent generation. The architecture integrates View-Temporal Attention (VTA) to capture cross-view spatial dependencies and temporal dynamics, together with Multi-view Rotary Position Encoding (RoPE) to encode positional relationships across views and time. WorldRec 2.0 reconstructs static and dynamic 3D Gaussians from multi-view, multi-temporal observations in a single forward pass and uses their renderings as explicit scene context to preserve geometry and appearance during long-horizon generation. Conditioning inputs are fused through VAE latent concatenation in a shared representation space. We incorporate Gated Linear Attention (GLA) into the temporal layers of VTA for efficient streaming inference, reducing computation and KV-cache memory during multi-view rollouts. A physical-interaction data pipeline covers diverse interactions and long-tailed events. From Pre-train onward, these data are combined with large-scale videos and heterogeneous visual modalities to learn visual, geometric, and dynamic priors. During Mid-train, training across diverse view configurations further improves the model's ability to capture how actions and object states shape subsequent scene evolution. After distillation-based Post-train, we apply the model to closed-loop simulation, closed-loop training, and World Action Model (WAM) applications. These applications generate physically consistent future states and support world-state prediction, policy learning, and action generation for Physical AI.
Figures & tables
Figure 1: Xiaomi Auto World Model 2.0 supports unified, streaming causal world modeling across general video generation, autonomous driving, and embodied robotics. Examples include text-to-video and image-to-video generation, controllable prediction across varying views, and action-conditioned rollouts for embodied manipulation. The resulting world model can serve as a simulator for both training and evaluation, and as a component in a WAM.
Method
Generation Duration
Camera views
Control
Domain
Applications
Real-time
LingBotWorld Team et al. (2026b)
Minutes
1V
Cam.,action
General
Interact.
✓
DreamXWorld Team et al. (2026a)
Minutes
1V
Cam.,action
General
Interact.
✓
CausalWM Xu et al. (2026e)
Seconds
1–4V
Action
Robot
Sim.
–
HorizonDrive Zhang et al. (2026)
Minutes
1V
Layout
Driving
Sim.
–
OmniDreams Basant et al. (2026)
Minutes
1/4/7V
Layout
Driving
Sim.,policy
✓
X-World Zheng et al. (2026)
∼ 24 s
7V
Layout
Driving
Sim.
✓
Table 1: Comparison of recent world models in terms of generation horizon, camera number, control, domain, downstream usage, and real-time. Cam. denotes camera pose; action denotes action sequences (including robot action sequences and WASD inputs); Interact. denotes open-interaction world models; Sim. denotes closed-loop simulation or training; policy indicates whether future action sequences are predicted.
Figure 2: Overview of our data pipeline. We integrate three data categories: general data, physical world driving and robotics data, and physical causal data. The collected videos undergo basic filtering, scoring, and captioning. We further prepare world-control conditions for the corresponding data subsets, including layouts, and scene contexts, to support causal world model training.
Figure 3: Overview of the dataset distribution. We report the sample count distribution, caption token count distribution, and scene composition of the general and physical-world autonomous-driving data. Robotics data are omitted because they are predominantly collected in a small number of indoor settings.
Figure 4: Architecture and training pipeline overview. The Causal World Model uses VTA and GLA to retain an efficient DiT backbone under multi-view settings. To handle diverse conditions, we adopt a unified condition injection design, enabling training across multiple tasks and datasets. During post-training, we introduce long-video rollout to improve long-horizon generation stability and Random Gradient Sampling to reduce memory usage. Overall, the model is trained in three stages to achieve efficient and controllable streaming generation.
Figure 5: WorldRec 2.0: point-aligned interaction and explicit 3D scene memory. Multi-camera and multi-temporal observations are aligned in 3D, exchanged through static-scene and dynamic-instance registers, decoded into static and dynamic Gaussians, and rendered under a target camera pose and time.
Data & Stage
Res.
Views
General Data
Physical World Data
Physical Causal Data
Full
HQ
Driving
Robot
Engine
LVM
Pub. Media
Data samples
–
–
3.6M
0.5M
5.4M
3.6M
37K
29K
12K
Pre-train
480p
1V
✓
✓
✓
✓
✓
✓
✓
Mid-train
480p,540p
1-7V
–
✓
✓
✓
✓
✓
✓
Post-train
480p,540p
1-7V
–
–
✓
✓
–
–
–
Table 2: Multi-stage training curriculum. In General Data, Full denotes the complete dataset, and HQ denotes the high-quality subset selected using the scoring scheme in Chapter 2. In Physical Causal Data, Pub. Media refers to counterfactual video data retrieved from public datasets and open-source materials (e.g., vehicle-collision clips).
Attention
SA/layer (ms) ↓
DiT/chunk (s) ↓
Latent FPS ↑
FID ↓
FSA
60.14
6.481
0.309
–
VTA
23.17 (2.60 × )
2.299 (2.82 × )
0.870 (+182%)
37.24
GLA( Ours )
8.412 (7.15 × )
1.742 (3.72 × )
1.148 (+272%)
38.08 (+2.3%)
Table 3: Efficiency and quality of three DiT attention designs. FSA denotes native full softmax attention and flattens all the view tokens; VTA denotes view-temporal decoupled attention with softmax attention; and GLA uses the VTA with GLA in temporal attention branch. SA/layer reports the mean CUDA-event latency (per DiT forward pass) of a representative attention layer used by each design: full spatiotemporal softmax attention for FSA , temporal-branch softmax attention for VTA , and temporal-branch gated linear attention for GLA . DiT/chunk is the causal DiT latency per chunk generation. Parenthesized efficiency values are speedups relative to FSA; the parenthesized FID change compares GLA with VTA.
Model
AV-Quality ↑
SC ↑
BC ↑
MS ↑
AQ ↑
IQ ↑
OC ↑
IS ↑
IB ↑
Wan2.2-5B Wan et al. (2025)
76.1
90.4
93.7
97.1
47.5
67.4
21.8
95.4
95.8
Wan2.1-14B Wan et al. (2025)
77.2
91.7
94.3
98.4
48.1
68.8
21.7
97.0
97.3
Wan2.2-14B Wan et al. (2025)
77.0
91.4
94.1
98.6
47.3
68.1
21.9
97.2
97.5
DreamXWorld Team et al. (2026a)
76.1
91.9
93.7
97.3
47.5
64.4
21.7
95.7
96.7
HunyuanVideo1.5 Wu et al. (2025a)
77.4
91.5
92.5
99.3
48.8
69.0
22.3
97.9
98.1
LingBotWorld Team et al. (2026b)
77.3
91.2
94.4
98.8
48.3
69.1
21.8
97.3
97.8
Table 4: Quality evaluation on the Autonomous Vehicle subset of PAI-Bench-G. AV-Quality denotes the average score; SC, BC, MS, AQ, IQ, OC, IS, and IB denote Subject Consistency, Background Consistency, Motion Smoothness, Aesthetic Quality, Imaging Quality, Overall Consistency, Image-to-Video Subject Consistency, and Image-to-Video Background Consistency, respectively. Bold and underlined values denote the best and second-best results per column.
Model
Robot-Quality ↑
SC ↑
BC ↑
MS ↑
AQ ↑
IQ ↑
OC ↑
IS ↑
IB ↑
Wan2.2-5B Wan et al. (2025)
74.3
89.0
91.0
98.0
46.1
67.8
19.2
90.3
92.8
Wan2.1-14B Wan et al. (2025)
75.2
87.9
90.3
98.5
48.8
71.7
19.0
91.5
93.7
Wan2.2-14B Wan et al. (2025)
75.4
89.4
90.6
98.4
48.1
70.6
19.4
92.3
94.3
DreamXWorld Team et al. (2026a)
74.8
90.3
92.2
97.3
45.6
66.2
18.9
92.4
95.8
HunyuanVideo1.5 Wu et al. (2025a)
76.9
89.0
90.6
99.2
51.0
70.1
19.6
97.9
97.8
LingBotWorld Team et al. (2026b)
77.0
91.8
93.2
99.0
50.0
66.8
19.7
97.8
98.0
Table 5: Quality evaluation on the Robot subset of PAI-Bench-G using the same metrics as Tab. 4 . Robot-Quality denotes the average score.
Figure 6: Qualitative results for long-tail corner cases. For compact visualization, we show the Front Wide, Rear, and Front Tele views selected from the full seven-view output. Rows present representative frames from 0 to 9 seconds. Left: Long-Tail Road Participants , where a horse appears in an urban driving scene. Right: Long-Tail Weather Conditions , showing nighttime driving on a snow-covered road.
Figure 7: Qualitative results for physically causal events. Three representative views are shown from the full seven-view output, with rows spanning 0 to 9 seconds. Left: Environmental Hazard Response , where rocks enter a mountain road and the ego vehicle performs an evasive maneuver. Right: Vehicle Collision Dynamics , where the ego vehicle progressively approaches and rear-ends the vehicle ahead.
Method
2011_09_26_drive_0002_sync
2011_10_03_drive_0047_sync
Reported Dataset Avg.
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
Cosmos-4B Agarwal et al. (2025)
0.138
84.0
0.155
77.5
0.155
77.1
Cosmos-12B Agarwal et al. (2025)
0.131
84.5
0.149
77.6
0.154
77.5
Copy Last (VGGT) Wang et al. (2025a)
0.112
89.6
0.140
84.4
0.137
83.7
Aether Zhu et al. (2025)
0.074
94.2
0.113
87.4
0.096
87.9
DINO-Foresight Karypidis et al. (2026)
0.063
96.5
0.076
90.9
0.082
91.6
Table 6: Depth prediction comparison on two driving sequences and the reported dataset average. AbsRel is lower-is-better, while δ1 (%) is higher-is-better.
Figure 8: Qualitative results for scene perception. Top: RGB frames and corresponding depth predictions. Bottom: RGB frames and referring segmentation masks for target objects. Columns show temporal progression of each sequence.
Method
Camera Controllability Score ↑
WorldScore
Combined
Average
CogVideoX-I2V Yang et al. (2025c)
38.27
26.72
32.50
Hunyuan 1.5 Wu et al. (2025a)
40.15
28.71
34.43
LingBot World Team et al. (2026b)
42.06
35.97
39.02
FantasyWorld-1.0 Dai et al. (2026)
81.45
44.28
62.87
DreamXWorld Team et al. (2026a)
81.23
43.01
62.12
Table 7: Camera-pose-controlled video generation evaluated using the camera controllability metric introduced by WorldScore. WorldScore denotes the static-world subset of the WorldScore evaluation set, while Combined denotes the evaluation set formed by merging TartanAir and RealCam-Vid. Higher scores indicate more accurate adherence to the specified camera trajectories.
Figure 9: Qualitative results for camera-pose-controlled generation. (a) Single-view camera control, with the prescribed camera trajectory shown on the left and representative generated frames on the right. (b) Multi-view camera control for the head, left, and right cameras, showing their trajectories and the corresponding synchronized observations. The color progression indicates the temporal order from start to end.
Figure 10: Qualitative results on internal driving data: 7-view long-video generation (45 s). Our model preserves fine-grained structures ( e.g. , lane markings and traffic signs) over long horizons.
Figure 11: Qualitative results on internal driving data: generation under novel trajectories. Our model remains robust under challenging, out-of-distribution trajectories.
Method
Views
Frames
FID ↓
FVD ↓
Short-video Generation
Vista Gao et al. (2024b)
1
25
6.90
89.40
DriveDreamer Wang et al. (2024a)
6
8
14.90
340.80
MagicDrive-V2 Gao et al. (2024a)
6
16
20.91
94.84
Ours
1
21
12.09
135.13
6
21
19.62
171.73
Table 8: Video generation results under short-video and long-video settings. FID and FVD are lower-is-better.
Figure 12: Qualitative visualization of world control (top) and our generated multi-view results (bottom). The world control is rendered from the robot URDF driven by the action sequence.
Method
TWB-Score
Traj. Acc.
Norm. PSNR
VQA Cons.
Flow Score
Bg. Cons.
Photo. Smooth.
DreamDojo Gao et al. (2026)
51.72
16.61
59.78
47.82
6.52
89.29
15.86
Fysiverse-Video Wang et al. (2026c)
63.62
37.40
68.90
63.32
22.16
75.36
35.63
WoVR_Plus Jiang et al. (2026b)
65.39
45.27
74.62
69.11
24.85
75.19
32.90
BWM Team and others (2026)
65.54
41.97
73.43
65.94
20.48
70.29
33.95
CausalWM Xu et al. (2026e)
66.04
42.98
72.41
70.03
22.99
75.53
30.27
Ours
64.79
47.25
73.44
69.31
23.76
75.57
34.16
Table 9: TriWorldBench results on TWB-Score and the metrics where our method ranks first or second among the compared methods. Bold and underlined scores denote first and second place. All metrics are higher-is-better.
Figure 13: Architecture of the closed-loop training strategy based on the causal reasoning world model simulator. The strategy comprises three core components: an end-to-end driving model, a reinforcement learning (RL) post-training module, and a causal reasoning world model simulator. Solid lines denote data flow; dashed lines denote gradient propagation. The proposed strategy fine-tunes the action generation model via post-training.
Method
Safety
Efficiency and Others
CR (%) ↓
SCR (%) ↑
OR (%) ↓
TTC (s) ↑
EP (%) ↑
HDR (%) ↑
DrivoR-E2E Kirby et al. (2026)
1.4
80.3
0.0
0.504
81.0
99.96
ST-JEPA-E2E Anonymous (2026)
1.2
95.2
0.0
0.614
84.7
99.93
E2E-WM-RL (Ours)
0.4
96.3
0.0
0.740
84.5
99.99
Table 10: Test Results of Closed-Loop Training Strategy and End-to-End Baseline Methods. CR denotes collision rate, SCR denotes success rate, OR denotes off-road rate, TTC denotes minimum time-to-collision, EP denotes ego progress rate, and HDR denotes trajectory compliance region rate. The upward arrow ↑ indicates that higher values correspond to better performance, while the downward arrow ↓ indicates that lower values correspond to better performance.
Figure 14: Visualization Results of Testing Closed-Loop Training Strategy and End-to-End Policy for Unsignalized Intersection Scenarios. ST-JEPA-E2E is the state-of-the-art end-to-end baseline, and E2E-WM-RL is our RL-posttrained end-to-end policy. Subplots from left to right show the world-model scene control diagram, world-model generated scene with trajectory projections, and BEV visualization with trajectory projections.
Figure 15: Visualization Results of Closed-Loop Training Strategy and End-to-End Policy for Parked Vehicle Occupation Scenarios. ST-JEPA-E2E is the state-of-the-art end-to-end baseline, and E2E-WM-RL is our RL-posttrained end-to-end policy. Subplots from left to right show the world-model scene control diagram, world-model generated scene with trajectory projections, and BEV visualization with trajectory projections.
Figure 16: World-Action Model. The causal world model encodes multi-frame observations into a spatiotemporal world latent. The Action Model then combines this latent with multi-camera history and ego status to produce action tokens. These tokens are decoded into candidate future trajectories, which are scored by the PDM heads.
Figure 17: Qualitative trajectory comparison for WAM planning. Candidate trajectories produced by the Action Model conditioned on the causal world model are visualized in the ego-centric scene, showing how the causal world model affects trajectory selection and driving behavior.
Method
Training data
NC ↑
DAC ↑
TTC ↑
C ↑
EP ↑
PDMS ↑
Epona Zhang et al. (2025)
Trainval 103k
97.9
95.1
93.8
99.9
80.4
86.2
DriveLaW Xia et al. (2025)
Trainval 103k
99.0
97.1
96.7
100.0
81.3
89.1
DriveVLA-W0 Li et al. (2026a)
Trainval 103k
98.7
99.1
95.3
99.3
83.3
90.2
DriveFine Dang et al. (2026)
Trainval 103k
98.6
97.9
95.2
99.9
85.5
90.7
R2SE Dang et al. (2026)
Trainval 103k
99.0
97.9
96.4
100.0
86.8
91.6
iPad Guo et al. (2026)
Trainval 103k
98.6
98.3
94.9
100.0
88.0
91.7
Table 11: WAM planning results on NAVSIM v1.1 navtest. External values are source-reported; dashes indicate unreported or pending submetrics. All metrics are higher-is-better.
Figure 18: Qualitative results on Waymo and nuScenes. WorldRec 2.0 produces geometry-consistent RGB renderings from unseen camera views and improves the reconstruction of dynamic regions.
Figure 19: Qualitative 7V reconstruction on the MiData.
Figure 20: Qualitative novel view results on the MiData.
Method
Dyn. PSNR ↑
Dyn. SSIM ↑
Dyn. D-RMSE ↓
Full PSNR ↑
Full SSIM ↑
Full D-RMSE ↓
EmerNeRF Yang et al. (2024a)
17.79
0.255
40.88
24.51
0.738
33.99
3DGS Kerbl et al. (2023)
17.13
0.267
13.88
25.13
0.741
19.68
PVG Chen et al. (2026c)
15.51
0.128
15.91
22.38
0.661
13.01
DeformableGS Yang et al. (2024b)
17.10
0.266
12.14
25.29
0.761
14.79
LGM Tang et al. (2024)
19.58
0.443
9.43
23.59
0.691
8.02
GS-LRM Zhang et al. (2024)
20.02
0.520
9.95
25.18
0.753
7.94
Table 13: Novel-view reconstruction on Waymo. Dynamic metrics are evaluated inside dynamic-object regions; full metrics use the complete image.
Streaming world models predict future visual states online while maintaining physically coherent dynamics over long horizons. However, their rollouts often violate basic physical constraints. A common approach distills pretrained bidirectional DiTs into few-step causal generators. However, this paradigm suffers from two fundamental limitations: generic bidirectional teachers acquire limited physical priors from visually oriented pretraining, and the limited priors suffer further loss during bidirectional-to-causal distillation. We present PhyS, a three-stage framework for distilling physical priors into streaming world models. To acquire physical priors from real-world interactions, we construct PhyS-120K, a dataset of 120K real-world physical-interaction videos spanning rigid-body dynamics, soft-body deformation, fluid phenomena, and phase transitions. Each video is annotated with structured descriptions of object properties and causal state transitions. Physics-aware supervised fine-tuning injects the physical priors into a bidirectional 14B DiT teacher, which we then distill into a lightweight 1.3B causal DiT for few-step autoregressive streaming generation. Finally, we use online reinforcement learning to incentivize the distilled model to generate physically plausible rollouts and further propose Temporal Credit Routing (TCR) to address temporal credit assignment. TCR evaluates physical consistency over overlapping temporal windows and routes the resulting group-relative advantages to temporally aligned denoising actions. On PhysicsIQ, PhyS improves the Wan2.1-14B teacher by 18.2% and the Self Forcing, Rolling Forcing, and Causal Forcing by 23.7%, 14.8%, and 31.4%, respectively. Results also improve the physics-aware video benchmarks VideoPhy, VideoPhy2, and PhyGenBench. The dataset, code, and more sample videos are available on our Project Page.
Liangliang Zhao, Junying Wang, Danni Yang +5
Fudan University, Shanghai, China · Shanghai AI Laboratory, Shanghai, China
This report presents a unified technical system addressing the two core capabilities of world models for autonomous driving: world representation and world generation. For world representation, we propose WorldRec, a feed-forward reconstruction architecture driven by sparse scene queries. WorldRec initializes structured queries in 3D space, leveraging them to aggregate cross-view, cross-temporal features, thereby naturally enforcing spatial consistency across frames and yielding compact yet high-fidelity 3D Gaussian scene representations. For world generation, we propose WorldGen, a two-stage training framework of bidirectional pretraining followed by causal fine-tuning through three progressive stages (Teacher Forcing, ODE distillation, and DMD), enabling high-quality online causal video generation in as few as 4 denoising steps. Building on both modules, we further introduce the JWM, which deeply integrates WorldRec and WorldGen to achieve synergistic gains in generation stability, cross-frame consistency, and visual fidelity, providing a solid foundation for closed-loop simulation, data synthesis, and end-to-end training in autonomous driving.
We introduce Kairos, a regret-aware native world-action model stack for Physical AI. Kairos is motivated by the view that a physical world model should not aim to fully simulate all future pixels, but should learn and maintain the information most relevant to embodiment control: object state, spatial relations, contact conditions, task progress, action consequences, failure boundaries, and deployment uncertainty. Kairos establishes three model-side prerequisites toward this goal. First, it learns control-relevant information through a Cross-Embodiment Data Curriculum, which organizes open-world videos, human behavioral data, and robot interactions into an intervention-strength progression from passive physical observation to intentional behavior and embodied action grounding. Second, it maintains control-sufficient states through a unified understanding, generation, and prediction architecture equipped with Hybrid Linear Temporal Attention, where local, mid-range, and global temporal pathways support multi-timescale state maintenance under efficient inference. Third, it deploys these states through a Deployment-Aware System Co-Design, treating latency, memory footprint, and hardware compatibility as first-order constraints for future observation, action, and feedback loops. Experiments on embodied world-model benchmarks, world-action benchmarks, long-horizon generation, and inference-efficiency evaluation show that Kairos achieves superior performance while offering a favorable efficiency to capability trade-off.