HiDream-O1-Embodied Tops RoboColiseum Robustness Benchmark, Unifies Multimodal World Models

World models are undergoing a critical shift. While models like Genie, Sora, and WorldLabs have advanced visual simulation, they remain statistical pixel predictors lacking physical and causal understanding. Embodied world models aim to align directly with physical rules, converting video generation into action control — predicting how a thrown ball falls, whether liquid spills when a robot lifts a cup, or how much force to apply when grasping a raw egg.

Global Competition and HiDream's Entry

Major players including Google, NVIDIA, and Tesla are investing heavily in simulation platforms and robot foundation models. Chinese companies are also competing aggressively. Today, HiDream.ai (智象未来) officially released its embodied world model HiDream-O1-Embodied . Built on a native multimodal philosophy, the model strengthens robotic physical perception and dynamic prediction, enabling higher-level physical interaction. Its release marks HiDream's progress toward a unified world model foundation that connects image, video, 3D, and action modalities across the full understand–reason–execute pipeline.

Benchmark Leadership: RoboColiseum Robustness

Upon release, HiDream-O1-Embodied was evaluated on RoboColiseum , a normalized embodied intelligence simulation benchmark featuring high-fidelity environments and 78 tasks across four dimensions: instruction following, spatial understanding, robustness (perturbation adaptation), and general manipulation. The platform attracts dozens of heavyweight models globally and provides reproducible, real-time rankings.

Robustness is widely considered the most challenging dimension. It tests model stability under varied backgrounds, lighting, materials, robot initial states, camera positions, image quality, and paraphrased instructions — essentially evaluating performance in non-ideal, real-world conditions rather than laboratory "greenhouse" settings.

HiDream-O1-Embodied scored 0.692 on the Robustness sub-leaderboard, securing the top position. A demonstration video of the "pick block shape" task (e.g., "right arm pick up the triangular prism on the table") shows the model identifying the target among clutter, estimating its 3D pose in real time, and planning a stable grasp trajectory despite deliberate disturbances such as lighting flicker, partial occlusion, and texture confusion.

Three Core Technical Innovations

Language Understanding — Beyond Keyword Matching

Traditional robots often fail when instructions are rephrased (e.g., "bring me the cup" vs. "hand me the cup"). HiDream-O1-Embodied covers a diverse equivalence space of verbs, sentence structures, and expressions, locking onto user intent regardless of phrasing.

Visual Perception — Multi-View Collaboration

Real-world visual channels are imperfect: camera drift, calibration decay, occlusions. The model fuses information from multiple viewpoints so that channels complement each other. When one view degrades, others sustain scene understanding and task execution, shifting the system from "single-point failure" to "graceful degradation."

High Fault Tolerance — Learning Stability in Imperfection

Most models train on pristine data. HiDream-O1-Embodied actively introduces non-ideal conditions during training — lighting changes, image corruption, occlusions, signal noise — forcing the model to make reliable judgments from limited cues. This yields robustness across diverse visual anomalies, ensuring consistent task completion even when conditions are far from perfect.

Model + Data: "Real Base + Generative Augmentation" Paradigm

This fault tolerance stems from a deeper data strategy. High-quality embodied data is scarce and decisive. HiDream employs a dual-driven "model + data" approach where the model actively participates in data creation. Collaborating with Noitom (诺亦腾), HiDream uses high-precision human motion capture as a real base, then leverages its native multimodal capability to achieve 100x fine-grained data augmentation. From a single real motion sample, the model generates variant videos that strictly preserve physical constraints while varying backgrounds, lighting, object shapes, and other factors.

The model acts as both examinee and examiner: it generates targeted training samples based on its own needs, creating a self-reinforcing data–model flywheel that ultimately produces extraordinary robustness under strong real-world perturbations.

Completing the Native Multimodal World Model Matrix

Less than a month earlier, HiDream released the interactive world model HiDream-O1-World , which topped the WBench Navi leaderboard with an average score of 80.9 . Interactive world models handle "understanding and reasoning" in digital spaces (spatial, temporal, motion, object relations), while embodied world models handle "operation and execution" in the physical world. Together they form complementary pillars for a native multimodal foundation.

As founder and CEO Dr. Mei Tao stated, next-generation large model competition hinges not on single-modality scaling but on the transition to natively unified multimodality. From HiDream-O1-Image to HiDream-O1-World to HiDream-O1-Embodied, HiDream is steadily building a model family covering visual, interactive, and embodied world models — demonstrating both cross-domain deployment of native multimodal technology and strong endogenous evolution capability.

The core of embodied intelligence is enabling AI to truly enter the real world, gaining continuous learning and adaptation in uncertain physical environments. From symbolic reasoning to digital large models to embodied interaction, AGI is completing a key evolution from virtual to real. AI is moving beyond "understanding the world" to "acting in the world, growing through interaction." The next world model competition will unfold in real physical scenarios, and Chinese enterprises intend to be full participants.