How DeepCybo’s Z‑WM Dominated WorldArena Track 2 with a 30.5‑Point Lead

One‑Year Milestone and Competition Success

In May 2026 DeepCybo marked its first anniversary and, shortly before that, its Z‑WM model achieved an 88.5‑point score on the WorldArena Track 2 Data Engine, leading the second‑place team by 30.5 points. The same model also placed eighth on the Track 1 overall leaderboard using only language‑driven input, surpassing many models that combine language and action.

WorldArena Evaluation Criteria

WorldArena Track 2 requires a model to generate future synthetic video streams from instructions; these streams are fed directly into downstream robot policy networks and executed in a physical simulation for closed‑loop grasping. The final score reflects the improvement in robot task success rate contributed by the synthetic data.

Human‑First‑Perspective Data Pipeline (ICDC)

DeepCybo’s strategy centers on “human‑first‑perspective” (egocentric) data, captured through the ICDC situational data acquisition system. Unlike tele‑operation or pure simulation data, ICDC records the causal context of actions—observations, judgments, and interactions—producing structured knowledge of spatial relations, object properties, and physical logic. This pipeline has yielded the DeepAct dataset, a multimodal collection of hundreds of thousands of hours of egocentric recordings covering diverse physical interactions.

Through the Egocentric2Embodiment conversion pipeline, raw first‑person video is transformed into structured supervision containing spatiotemporal relationships, object attributes, force information, and reasoning traces, enabling embodied base models to learn from real‑world experience.

Base Model: PhysBrain 1.0

In March 2026 DeepCybo released PhysBrain 1.0, the first domestic embodied‑general‑intelligence (E‑AGI) base model pretrained on zero‑real‑robot trajectories. PhysBrain 1.0 rests on three original technologies:

PhysBrain Data Pipeline : Scales extraction of implicit physical experience from egocentric video into structured supervision.

TwinBrainVLA Dual‑Brain Architecture : A frozen left brain retains general semantic understanding while a trainable right brain focuses on fine‑grained action strategies, addressing catastrophic forgetting.

LangForce Training Strategy : A Bayesian decomposition maximizes mutual information between actions and instructions before motion generation, ensuring the robot “listens before acting”.

With only 3,000 hours of high‑density egocentric data for pretraining, PhysBrain 1.0 achieved 80.2 % success on SimplerEnv WidowX and 91.3 % on Google Robot, surpassing industry baselines such as Pi0.5 (57.1 %). The model also displayed autonomous error‑correction and flexible execution strategies not present in the training data.

Capability Extensions

To further strengthen the base model, DeepCybo introduced plug‑and‑play modules:

Euclid’s Gift : Uses Euclidean geometry problems as proxy tasks to inject strong spatial‑reasoning priors, achieving top ranks on VSI‑Bench and MindCube with zero‑shot transfer.

3D‑Mix : A lightweight gated module that adds three‑dimensional perception to VLA variants, improving out‑of‑distribution performance by an average of 7 % absolute.

IntentVLA : Maps recent visual history to short‑term intent beliefs, reducing execution ambiguity in partially observable scenes and boosting stability across several leaderboards.

World Model and Strategy Layer

Building on the mature base, DeepCybo added a closed‑loop world‑model and strategy layer:

EA‑WM addresses physical realism of synthetic data by rendering kinematic information into camera‑aligned visual streams (KVAF) and employing an Event‑Driven Latent Sensing (EDLS) mechanism that focuses on contact moments, ensuring generated videos obey physical laws.

STARRY converts high‑quality synthetic data into precise manipulation policies. Its Geometry‑Aware Selective Attention Modulation (GASAM) directs the policy network’s attention to critical action regions, markedly improving fine‑grained operation accuracy.

The full loop—EA‑WM improving data realism, STARRY turning data into robot policies, and WorldArena Track 2 validating end‑to‑end task success—demonstrates the effectiveness of DeepCybo’s pipeline.

Hardware Platform: Prime Series

DeepCybo’s hardware complements its software stack. The Prime robot is the world’s first full‑size humanoid capable of autonomous standing after power loss (173 cm tall, 72 DOF). It embodies the “human data → humanoid data” mapping, allowing the model’s physical intuition to translate into precise control. Variant models Prime U and Prime Lite target real‑task execution and educational scenarios respectively.

Impact and Outlook

The Track 2 victory signals a shift in world‑model evaluation: scores now depend on robot task completion rather than video frame quality. While many teams excel at generating high‑fidelity videos (Track 1), few produce data that can directly train robots. DeepCybo’s achievement demonstrates that closing the loop from egocentric data to embodied policies is feasible.

If high‑fidelity synthetic data generation matures, robot data‑collection costs could drop dramatically, accelerating embodied‑AI commercialization. Nonetheless, the current results rely on simulation‑based closed‑loop tests; generalization to the real world remains an open research question. Early real‑world validation on the ARX R5 dual‑arm robot shows success rates rising from 42.5 % to 70.8 % when using STARRY‑generated policies.

Overall, DeepCybo’s year‑long technical roadmap—from data paradigm to base model, spatial intelligence, world model, and hardware—forms a cohesive system that may become a long‑term competitive moat in embodied AI.