Video-based world models offer a powerful paradigm for embodied simulation and planning, yet state-of-the-art models often generate physically implausible manipulations - such as object penetration and anti-gravity motion - due to training on generic visual data and likelihood-based objectives that ignore physical laws. We present ABot-PhysWorld, a 14B Diffusion Transformer model that generates visually realistic, physically plausible, and action-controllable videos. Built on a curated dataset of three million manipulation clips with physics-aware annotation, it uses a novel DPO-based post-training framework with decoupled discriminators to suppress unphysical behaviors while preserving visual quality. A parallel context block enables precise spatial action injection for cross-embodiment control. To better evaluate generalization, we introduce EZSbench, the first training-independent embodied zero-shot benchmark combining real and synthetic unseen robot-task-scene combinations. It employs a decoupled protocol to separately assess physical realism and action alignment. ABot-PhysWorld achieves new state-of-the-art performance on PBench and EZSbench, surpassing Veo 3.1 and Sora v2 Pro in physical plausibility and trajectory consistency. We will release EZSbench to promote standardized evaluation in embodied video generation.
Figures & tables
Figure 1
Figure 1: Overview of the data curation pipeline. (a) shows the multi-stage filtering and balancing flow from raw aggregation ( ∼ 3M clips) to training-ready splits (SFT, RL, and A2V data). (b) Task-aware quota allocation: head tasks are capped at 8–15%, body tasks are uniformly sampled at 40–50%, and long-tail tasks are fully preserved to maximize task diversity. (c) Dataset and robot type distribution: the left ring shows the original composition and the right ring shows the rebalanced result after hierarchical sampling. (d) Physics-aware video captioning pipeline: a perception module (Qwen3-VL 32B) extracts structured physical attributes, followed by a writing module (Qwen3 32B FP8) that generates four-phase captions covering scene setup, action detail, state transition, and camera summary.
Figure 2: Two-stage training pipeline. Stage 1 : SFT on the DiT to predict future frames from observations and instructions. Stage 2 : generate N candidates, score via physics checklist, and apply DPO via LoRA on frozen DiT weights.
Figure 3: Construction pipeline of the EZSbench. Top : dual-source image augmentation—Branch 1 generates synthetic initial observations via text-to-image (Nano Banana) by varying robot morphology, scene, task, and viewpoint; Branch 2 applies VLM-guided background editing to real-world images while preserving foreground interactions. Down : three-stage dense description synthesis—visual anchoring grounds the scene layout and object coordinates, action simulation infers kinematically compliant trajectories with micro-physical interactions, and narrative synthesis produces a documentary-style caption integrating initial state, trajectory, and final state.
Model
AQ
BC
IQ
MS
OC
SC
I2VB
I2VS
Quality Score
Domain Score
Avg.
Wan 2.5
0.5477
0.8985
0.6458
0.9623
0.2190
0.8784
0.9438
0.9428
0.7548
0.8644
0.8096
GigaWorld-0
0.4757
0.9219
0.6506
0.9908
0.1944
0.9111
0.9673
0.9607
0.7591
0.8583
0.8087
Veo 3.1
0.5458
0.9216
0.7244
0.9712
0.2214
0.9146
0.9317
0.9614
0.7740
0.8350
0.8045
Wan2.1_14B
0.4723
0.9315
0.7118
0.9917
0.1921
0.9185
0.9745
0.9451
0.7672
0.8391
0.8032
WoW-wan 14B
0.4664
0.9295
0.7027
0.9858
0.1941
0.9149
0.9613
0.9292
0.7605
0.8301
0.7953
Cosmos-Predict 2.5
0.4897
0.9166
0.7405
0.9906
0.1911
0.8973
0.9304
0.9030
0.7574
0.8021
0.7797
Table 1: Quantitative comparison on the PAI-Bench robot domain subset.
Figure 4: Architecture of the action-conditioned video generation model. We selectively duplicate DiT blocks as parallel context blocks to process action maps, and fuse their outputs residually into the main DiT.
Model
AQ
BC
IQ
MS
OC
SC
I2VB
I2VS
Quality Score
Domain Score
Avg.
WoW-wan 14B
0.4764
0.9412
0.7514
0.9922
0.1236
0.9347
0.9495
0.9178
0.7609
0.7951
0.7780
GigaWorld-0
0.4466
0.8893
0.6565
0.9889
0.1222
0.8678
0.9325
0.9139
0.7272
0.7826
0.7549
Cosmos-Predict 2.5
0.4148
0.8835
0.6810
0.9878
0.0970
0.8366
0.9054
0.8653
0.7089
0.7698
0.7394
UnifoLM-WMA-0
0.4861
0.9575
0.7390
0.9925
0.1046
0.9452
0.8421
0.8170
0.7355
0.5232
0.6294
Our Model
0.4789
0.9494
0.7483
0.9910
0.1301
0.9442
0.9680
0.9453
0.7694
0.8366
0.8030
Table 2: Quantitative comparison on EZSbench.
Figure 5: Qualitative comparison on PAI-Bench.
Model
PSNR
SSIM
Traj. Consis.
Enerverse-AC
20.42
0.7542
0.8157
Gen-Sim
18.05
0.7413
0.6195
Ours
21.09
0.8126
0.8522
Table 3: Quantitative results on action-conditioned generation.
Figure 10
Figure 6: Vision-action alignment verification. Calibrated action signals are rendered as semi-transparent action maps and overlaid onto video frames. Samples with spatial deviation between projected trajectories and observed robot motion are identified and removed.
Figure 7: Stage 1 captioning example 1. Structured perception output showing extracted physical attributes, object identities, and spatial relations.
Figure 8: Stage 1 captioning example 2. The perception module identifies robot morphology, object properties, and contact events from the video sequence.
Figure 9: Stage 1 captioning example 3. Spatial layout parsing and state transition detection across the manipulation trajectory.
Figure 10: Stage 2 captioning example 1. The writing module synthesizes a four-phase narrative covering scene setup, action detail, state transition, and camera summary from the structured perception output.
Figure 11: Stage 2 captioning example 2. Physics-grounded narrative synthesis capturing fine-grained manipulation dynamics and causal state transitions.
Figure 13: Case study on zero-shot test set. Our model handles diverse unseen manipulation tasks including multi-object attribute binding, deformable object manipulation with dual-arm coordination, articulated object interaction, rigid body placement, contact-intensive wiping, and object relocation—all with strict physical plausibility and spatiotemporal coherence.
Figure 14: Action-to-video qualitative comparison. Our method preserves object geometry and visual integrity during contact-intensive manipulation. Baselines (Genie-Envisioner, Enerverse-AC) produce object deformation, contactless grasping, and localization errors.
Video generation models have emerged as a promising paradigm for embodied world simulation. However, both general-domain video generators and robot-specific data fine-tuned models can still produce physically implausible manipulations, including discontinuous motion trajectories and inconsistent robot-object interactions, which limits their reliability as world simulators. Through extensive experiments, we find that such physical instability mainly arises from two factors: deformation of moving objects and implausible spatio-temporal correlations among interacting entities, particularly during contact. Building on this observation, we propose PhysisForcing, a scalable training framework that strengthens physical consistency by focusing supervision on physics-informative regions through joint optimization of pixel-level and semantic-level features. The framework consists of a pixel-level trajectory alignment loss, which supervises DiT features using reference point trajectories, and a semantic-level relational alignment loss, which aligns DiT features with inter-region relations extracted from a frozen video understanding encoder. Extensive experiments on R-Bench, PAI-Bench, and EZS-Bench show that PhysisForcing consistently improves embodied video generation over strong baselines, improving the Wan2.2-I2V-A14B and Cosmos3-Nano base models on R-Bench by 22.3% and 9.2% (7.1% and 3.7% over vanilla finetuning), with the Cosmos3-Nano variant attaining the best overall score. Beyond generation, as a world model under the WorldArena action-planner protocol it raises the closed-loop success rate from 16.0% to 24.0% and further improves downstream policy success, indicating that physically aligned video models yield stronger representations for robotic manipulation.
Recent advances in large-scale video world models have enabled increasingly realistic future prediction, raising the prospect of using generated videos as scalable supervision for robot learning. However, for embodied manipulation, perceptual realism alone is not sufficient: generated interactions must also be physically consistent and executable by robotic agents. Existing benchmarks provide valuable assessments of visual quality and physical plausibility, but they do not systematically evaluate whether predicted behaviors can be translated into executable actions that complete manipulation tasks. We introduce RoboWM-Bench, a manipulation-centric benchmark for embodiment-grounded evaluation of video world models. RoboWM-Bench converts generated human-hand and robotic manipulation videos into embodied action sequences and validates them through execution in physically grounded simulation environments. Built on real-to-sim scene reconstruction and diverse manipulation tasks, RoboWM-Bench enables standardized, reproducible, and scalable evaluation of physical executability. Using RoboWM-Bench, we evaluate state-of-the-art video world models and observe that visual plausibility and embodied executability are not always aligned. Our analysis highlights several recurring factors that affect execution performance, including spatial reasoning, contact prediction, and non-physical geometric distortions, particularly in complex and long-horizon interactions. These findings provide a more fine-grained view of current model capabilities and underscore the value of embodiment-aware evaluation for guiding physically grounded world modeling in robotic manipulation.
Feng Jiang, Yang Chen, Kyle Xu +8
Peking University · Tsinghua University · Lightwheel
We introduce Qwen-RobotWorld, a language-conditioned video world model for embodied intelligence. With natural language as a unified action interface, it predicts physically grounded future visual trajectories from current observations across robotic manipulation, autonomous driving, indoor navigation, and human-to-robot transfer. This unified formulation provides three promising application directions: synthetic data generation for policy training augmentation, scalable virtual environments for policy evaluation, and language-guided planning signals for downstream robot control. This is achieved through a three-part design: a) Double-Stream MMDiT with MLLM Action Encoding, where a 60-layer double-stream diffusion transformer couples frozen Qwen2.5-VL semantics with video-VAE latents through layer-wise joint attention; b) Embodied World Knowledge (EWK), an 8.6M video-text corpus (200M+ frames) with action-language mapping over 20+ embodiments and 500+ action categories; and c) General+Expert Progressive Curriculum, a two-stage training strategy that first learns general visual priors and then injects embodied specialization under a shared language interface. Extensive results show strong competitiveness: ranks 1st overall on EWMBench and DreamGen Bench, outperforms all open-source models on WorldModelBench and PBench. Additional zero-shot analyses on RoboTwin-IF benchmark further support robust generalization and multi-view consistency.