蚂蚁灵波沈宇军：区分模型能力和模型展现的能力 「触觉共识」将率先形成｜WAIC 2026

This article compiles key insights on physical AI and world models from Shen Yujun, Chief Scientist of Ant Lingbo, delivered at the WAIC 2026 World Model Summit. Key takeaways are as follows:

1. It is critical to distinguish between a model's intrinsic capability and its demonstrated performance. While current industry investments in model architecture and data infrastructure are valuable, over-prioritizing flashy demonstration outcomes has questionable merit. The biggest gap holding AI back from widespread adoption in the physical world is the lack of unified data standards, which are required for industry-wide scalable expansion.

2. A high-quality world model can be measured by three core metrics: First, its ability to predict counterfactual future states based on current conditions and actions, which demonstrates true understanding of the world. Second, its ability to retain invariant world representations and extract shared universal representations of the environment. Third, its ability to conduct high-temporal-efficiency inference; long-horizon prediction is a false demand for embodied models.

3. The industry will first reach consensus on the tactile modality, which will become a core required capability for physical AI-powered robots.

This article outlines the latest development trends in physical AI, offering key takeaways for brands positioning AI-related businesses:

1. AI deployment in physical scenarios is a major long-term industry direction, but the current lack of unified data standards and modality alignment norms means large-scale commercialization will take time. Brands should reserve sufficient room for technical iteration when rolling out new AI-related businesses, and avoid rushing into large-scale blind investments.

2. AI deployed in offline consumer scenarios frequently encounters unplanned emergent situations, such as service robots handling customers' non-standard requests. Brands should fully test an AI product's performance across a full range of extreme daily scenarios before launch.

3. The industry will first reach consensus on the tactile modality, and AI hardware with tactile interactive capabilities will emerge as a hot new product category. Brands can monitor technical iteration early and pre-position in this category to capture first-mover advantage.

This update on physical AI industry progress offers key takeaways for sellers positioning AI-related products or services:

1. World models and physical AI are still in the stage of rapid technical iteration, with no unified data or evaluation standards, and large-scale commercialization will take time. Sellers should avoid blindly following hype and overcommitting resources to prevent losses.

2. The core of successful AI commercialization is cost control, paired with the ability to handle unplanned emergent scenarios. When selecting AI products or launching AI-related services, sellers should prioritize calculating total deployment costs, and focus testing on performance in non-standard daily scenarios to improve user experience.

3. The industry will first reach technical consensus in the tactile field, and related tactile interactive AI products will see new growth opportunities. Sellers can pre-position in this track to capture first-mover advantage and growth dividends.

This presentation offers key insights for factories advancing digital transformation and developing AI-related production and R&D:

1. When deploying physical AI and embodied intelligence in factory settings, do not only judge by demo performance. Instead, focus on actual operational losses and total investment costs. The core of commercializing technology is economic viability, so cost is the primary criterion for deployment.

2. Currently the industry lacks unified data and evaluation standards, and the technology is still immature. When introducing AI technology to upgrade production or develop new products, factories should first run small-scale tests to verify performance before expanding gradually, to avoid losses from large-scale unproven investments.

3. Tactile perception will become a core required capability for all robots, including both industrial and service robots, and new tactile interaction requirements will be added to related product design and production. Factories can build up relevant technical reserves and adjust production lines in advance to capture new industry opportunities.

This presentation clarifies industry trends and core pain points in physical AI and world models, offering key insights for AI technology service providers:

1. The most pressing core pain point facing the industry today is the lack of unified data standards and cross-modality alignment, which blocks industry-wide scalable expansion. Service providers can develop new solutions focused on building unified data standards and enabling multi-modality alignment to tap into new market opportunities.

2. The two most prominent pain points for clients deploying AI are excessive costs and inability to handle unplanned emergent scenarios, as existing demos cannot cover random daily situations. Service providers can optimize model performance for handling emergent scenarios across different industry use cases, while helping clients control deployment costs to improve the competitiveness of their solutions.

3. The next phase of industry convergence will center on evaluation standards, and consensus will first emerge in the tactile field. Service providers can start R&D for tactile-related AI technology early to secure position in this emerging market.

This presentation reveals core demands and development directions of the physical AI industry, offering key insights for marketplace platforms positioning in the AI sector:

1. The most critical shared demand in the physical AI industry today is unified data standards and unified evaluation standards. Platforms can take the lead to work with leading industry players to co-develop these standards, build industry barriers, attract more relevant enterprises to join the platform, and enrich the platform ecosystem.

2. Many AI companies currently face common challenges with cost control and handling emergent scenarios during deployment. Platforms can launch targeted enablement services to help member companies solve these deployment pain points and improve user retention.

3. Consensus will first emerge in the tactile field, and related startups will see rapid growth. Platforms can launch targeted recruitment for tactile AI-related enterprises in advance to position for the next phase of track growth and capture first-mover advantage.

This presentation reflects the latest industry developments in the integration of world models and embodied intelligence, offering key insights for related researchers:

1. Industry has reached a consensus that AI integration into the physical world will become a foundational paradigm for the sector. The most critical unresolved industry issue is the lack of unified data standards, which are required for industry-wide scalable expansion, making this a highly important frontier research direction.

2. Industry has proposed three core evaluation metrics for world models, explicitly noting that long-horizon prediction is a false demand for embodied world models, with inference timeliness prioritized instead. It also notes that evaluation standards will be the focal point of industry convergence in the next phase, providing clear direction for related research.

3. Industry expects consensus to first emerge in the tactile field (the "tactile consensus"), with tactile explicitly identified as the core modality that distinguishes physical AI from digital world models. This is a new development at the industry level with high research value.