WAIC Viewpoint: The 'ChatGPT Moment' for Robots Could Arrive in as Little as Two Years

Overcoming the 'Three Walls' of Embodied AI

The development of embodied AI, particularly the concept of 'world models' that allow robots to simulate and understand physical reality, faces formidable challenges. Experts at the WAIC 2026 identified three primary 'walls' hindering progress: the data wall, the representation wall, and the closed-loop wall. The data wall refers to the immense gap between the data volume needed for training sophisticated physical AI models and what is currently available; it's estimated that physical world data is orders of magnitude less dense than linguistic data, requiring a 10,000-fold increase to match the scale of large language models. This scarcity necessitates innovative data collection strategies, ranging from low-cost, body-agnostic methods like UMI data to the insistence on high-fidelity real-machine data, and multi-layered 'data pyramid' approaches combining various sources.

The representation wall highlights the absence of a unified physical representation that can span diverse tasks, scenarios, and robot bodies, making it difficult for models to generalize. Finally, the closed-loop wall points to the high cost and slow feedback inherent in real-world trial-and-error, which severely limits the scalability of training. Addressing these fundamental issues is paramount for embodied AI to move beyond laboratory demonstrations and achieve widespread deployment.

The Evolving Landscape of Robot Intelligence Models

The architectural debate within embodied AI is intensifying, with discussions around the efficacy of Vision-Language-Action (VLA) models versus World Action Models (WAM). While some practitioners suggest that VLA models might be 'dead' in favor of WAM, others argue that VLA still holds promise, citing examples of cross-body migration capabilities. The consensus appears to be that both VLA and WAM approaches need continuous evolution and eventual convergence, potentially leading to a 'World Reasoning Action Model' (WRAM) that integrates comprehensive world understanding with sophisticated reasoning and action generation. This reflects a broader recognition that embodied intelligence requires more than just large-scale models; it demands a multi-layered agent system capable of complex cognition, perception-action loops, and underlying reactive systems, moving beyond simple control mechanisms.

Companies like Zhiyuan are actively developing foundational models (e.g., GO-2, Act2Goal) and distributed reinforcement learning systems (Genie Evolver) to bridge these architectural gaps. The shift towards more integrated and robust model architectures is seen as crucial for enabling robots to handle the complexities and uncertainties of real-world environments, moving closer to human-like intelligence in physical tasks.

Real-World Validation and the Path to a Standard

Significant progress is being made in validating embodied AI in real-world scenarios, demonstrating the potential for practical applications. Companies are showcasing impressive deployment successes, such as Zhiyuan robots achieving a 99.99% success rate in 3C production lines and Dyna Robotics' robots folding napkins with 99.4% accuracy in a commercial restaurant setting. These examples underscore a critical shift from 'accidental successes' in labs to repeatable, verifiable performance in actual business operations. Furthermore, advancements in training efficiency, like Zhiyuan's Genie Evolver system, which compresses training cycles to minutes by leveraging over 100 robots for online policy iteration, are accelerating development.

However, a major impediment to industry-wide progress is the lack of a standardized, fair, and unified evaluation benchmark for world models. Current metrics often focus on visual generation rather than the crucial physical causal reasoning abilities. To address this, China has launched Physical IQ Wuzhi, the country's first embodied native, multi-scenario, multi-body, multi-task physical intelligence evaluation platform. This platform aims to provide a common protocol for assessing advanced capabilities like dual-arm collaboration, long-range task planning, and tool use across diverse real-world tasks in retail, industrial, and home settings, fostering a more structured and competitive development environment for embodied AI.