cs.CVMar 13, 2026
SaveRoboStereo: Dual-Tower 4D Embodied World Models for Unified Policy Optimization
Organizations: Tsinghua University · X Square Robot · HKUST
Abstract
Scalable Embodied AI faces fundamental constraints due to prohibitive costs and safety risks of real-world interaction. While Embodied World Models (EWMs) offer promise through imagined rollouts, existing approaches suffer from geometric hallucinations and lack unified optimization frameworks for practical policy improvement. We introduce RoboStereo, a symmetric dual-tower 4D world model that employs bidirectional cross-modal enhancement to ensure spatiotemporal geometric consistency and alleviate physics hallucinations. Building upon this high-fidelity 4D simulator, we present the first unified framework for world-model-based policy optimization: (1) Test-Time Policy Augmentation (TTPA) for pre-execution verification, (2) Imitative-Evolutionary Policy Learning (IEPL) leveraging visual perceptual rewards to learn from expert demonstrations, and (3) Open-Exploration Policy Learning (OEPL) enabling autonomous skill discovery and self-correction. Comprehensive experiments demonstrate RoboStereo achieves state-of-the-art generation quality, with our unified framework delivering >97% average relative improvement on fine-grained manipulation tasks.
Figures & tables
Figure 1: (a) Qualitative comparison of RoboStereo against SOTA EWMs. (b) Quantitative comparison of unified policy optimization framework against traditional paradigms.
Figure 2: RoboStereo Architecture. Symmetric dual DiT towers (a) process RGB and XYZ pointmaps via bidirectional cross-attention for visual-geometric fusion (b) and a Gaussian head for flexible-viewpoint rendering. Dual-path action-conditioned timestep embedding mechanism (c) ensures precise frame-level trajectory control.
Figure 3: Illustration of Test-Time Policy Augmentation (TTPA).
Figure 4: Illustration of Imitative-Evolutionary Policy Learning (IEPL).
Figure 5: Illustration of Open-Exploration Policy Learning (OEPL).
Figure 6: Training procedures for IEPL and OEPL. IEPL (left) learns through visual imitation by matching policy rollouts with expert trajectories. OEPL (right) learns through open exploration guided by a discriminator-based reward model.
| Models | Visual Quality | Motion Quality | Content Consistency | ||||||
| Image | Aesthetic | JEPA | Dynamic | Flow | Motion | Subject | Background | Photometric | |
| Quality | Quality | Similarity | Degree | Score | Smoothness | Consistency | Consistency | Consistency | |
| GigaWorld | 0.483 | 0.384 | 0.429 | 0.650 | 0.308 | 0.771 | 0.727 | 0.825 | 0.172 |
| Genie | 0.222 | 0.319 | 0.327 | 0.666 | 0.083 | 0.687 | 0.762 | 0.875 | 0.197 |
| MIND-V | 0.319 | 0.397 | 0.452 | 0.552 | 0.267 | 0.756 | 0.803 | 0.902 | 0.240 |
| RoboMaster | 0.346 | 0.378 | 0.294 | 0.606 | 0.146 | 0.693 | 0.809 | 0.896 | 0.333 |
Table 1: Video evaluation results across visual quality, motion quality and content consistency dimensions. Best results are bolded , and second-best results are underlined .
Figure 7: Visualizations of the 4D Gaussian representations, RGB videos, and depth maps produced by RoboStereo. RoboStereo generates precise future trajectories conditioned on action instructions, exhibiting high visual fidelity and geometric consistency.
| Models | Physics Adherence | 3D Accuracy | Controllability | ||||
| Interaction Quality | Trajectory Acc. | Depth Acc. | Perspectivity | Instruction Follow. | Semantic Align. | Action Follow. | |
| GigaWorld | 0.523 | 0.152 | 0.611 | 0.753 | 0.599 | 0.844 | 0.109 |
| Genie | 0.198 | 0.067 | 0.844 | 0.513 | 0.200 | 0.848 | 0.011 |
| MIND-V | 0.570 | 0.137 | 0.705 | 0.788 | 0.593 | 0.865 | 0.079 |
| RoboMaster | 0.526 | 0.115 | 0.801 | 0.731 | 0.561 | 0.839 | 0.075 |
| Vidar | 0.520 | 0.185 | 0.775 | 0.748 | 0.573 | 0.856 | 0.039 |
Table 2: Video evaluation results across physics adherence, 3D accuracy and controllability dimensions. Best results are bolded , and second-best results are underlined .
| ID | Methods | Coffee | StackThree | Square | Mean |
| 0 | Base policy | 32.6 | 30.2 | 20.2 | 27.7 |
| 1 | TTPA | 46.7 | 46.9 | 36.2 | 43.3 |
| 2 | DPO | 32.6 | 38.2 | 23.3 | 31.4 |
| 3 | GRPO | 29.8 | 35.1 | 22.7 | 29.2 |
| 4 | IEPL | 47.8 | 44.1 | 37.4 | 43.1 |
| 5 | OEPL | 53.4 | 50.8 | 39.3 | 47.8 |
Table 3: Task success rates (%) under different optimization paradigms in the MimicGen simulation benchmark.
| Model Variant | Physics Adherence | 3D Accuracy | ||
| Interaction Quality | Trajectory Acc. | Depth Acc. | Perspectivity | |
| (a) Video tower only | 0.538 | 0.320 | 0.821 | 0.758 |
| (b) Dual-tower joint training | 0.570 | 0.342 | 0.881 | 0.797 |
| Model Variant | Controllability | Motion Quality | ||
| Instruction Follow. | Action Follow. | Flow Score | Motion Smoothness | |
| (c) Action-cond via channel concatenation | 0.622 | 0.061 | 0.238 | 0.703 |
Table 4: Ablation study conducted on the Bridge dataset. Results demonstrate that joint dual-tower training substantially improves 3D consistency in generated videos, while incorporating action conditioning via timestep embedding modulation yields markedly superior action control and motion quality. Best results are bolded .
| Method | Speed (FPS) |
| WoW [ 7 ] | 0.05 |
| MIND-V [ 48 ] | 0.38 |
| IRASim [ 51 ] | 0.53 |
| RoboMaster [ 10 ] | 0.63 |
| Ours (Video Tower) | 1.50 |
Figure 8: Inference speed comparison.
Figure 9: Visualization of fine-grained geometric reasoning enabled by IEPL on the Three_Piece_Assembly_D0 task. The base policy (top row) produces coarse actuation and fails to satisfy precise 3D geometric constraints. In contrast, the IEPL-optimized policy (bottom row), guided by dense visual imitation rewards from RoboStereo’s imagined trajectories, performs continuous micro-adjustments, achieving accurate part insertion.
Figure 10: Visualization of emergent self-correction via OEPL on the Square_D0 task. The base policy (top row) exhibits open-loop failure propagation: after an initial grasp failure, it erroneously treats the failed state as successful and blindly executes the subsequent transport trajectory. In contrast, the OEPL-optimized VLA policy (bottom row) autonomously detects the grasp slip and initiates a corrective re-grasping maneuver, achieving robust task completion.
Figure 11: Qualitative results of real-world deployment. The robotic arm is instructed to clear target cubes from a desktop with a prism serving as a distractor. The VLA jointly optimized via IEPL and OEPL exhibits emergent self-correction and fine-grained micro-adjustment capabilities. These critical recovery behaviors absent in the imitation learning baseline.
| Hyperparameter | Value |
| Optimizer | AdamW( ) |
| Learning rate | |
| Training batch size | 64 |
| Group size | 8 |
| Mini-batch size | 128 |
| Clip ratio | 0.20 |
Table 5: Hyperparameters for the GRPO algorithm.
| Model Dim | FFN Dim | Heads | Head Dim | Blocks | ||
| Self-Attn | Text Cross-Attn | Tower Cross-Attn | ||||
| 2048 | 8192 | 16 | 128 | 56 | 56 | 28 |
Table 6: Model Architecture Configuration
| Dimension | Metric | Computation Method |
| Visual Quality | Image Quality | No-reference frame sharpness via MUSIQ [ 13 ] ; averaged over all frames. |
| Aesthetic Quality | Per-frame aesthetic score via LAION Aesthetic Predictor [ 17 ] ; averaged over all frames. | |
| JEPA Similarity | MMD between V-JEPA [ V-jepa ] feature distributions of generated and reference videos using a second-order polynomial kernel; exponentiated for normalization. | |
| Motion Quality | Dynamic Degree | Top-5% optical flow magnitudes (RAFT [ 5 ] ) mapped via sigmoid to quantify salient motion intensity. |
| Flow Score | Mean optical flow magnitude across all pixels and frames via RAFT; reflects overall motion intensity. | |
| Motion Smoothness | SSIM between VFI-predicted [ 42 ] and real intermediate frames, weighted by motion magnitude to penalize static sequences. |
Table 7: Summary of video quality evaluation metrics in WorldArena [ 33 ] .