Multi-variety, small-batch, fast-changeover become the norm. Material states, positions, and working conditions vary, making exception handling and human-robot collaboration more valuable.

Population aging and skilled labor shortages intensify constraints. In 2025, China's total population reached 1.404 billion, with 23% aged 60+.

Industrial Embodied AI is the key path to "task-closed-loop automation", enabling systems to continuously complete perception, decision, execution, verification, and exception handling under process, quality, and safety constraints.

Online loop: Task input → perception & state estimation → task planning → policy & control → execution & feedback, ensuring stable task completion.

Offline loop: Data collection → governance → simulation training → evaluation & release → deployment feedback, enabling continuous improvement.

Currently, many industrial tasks still rely on rule-based orchestration and specialized models; large models are mainly used for task understanding and exception handling.

Upstream: from hardware components to "hardware base + intelligent capabilities + software infrastructure" composite layer, with deployment-closed-loop capability as the new standard.

Midstream: closest to customer value, with overlapping roles in carriers, scenario solutions, and integration; system ownership still converging.

Downstream: industrial clients define demand and acceptance criteria; repurchase and expansion drive commercialization.

Player types: carrier platform, scenario solution, integration delivery, model capability, software & infrastructure —协同 forming task-closed-loop.

Mature Logistics & Fixed-Station: clear task boundaries, quantifiable value, ready for replication. Reps: Hikrobot, UR.

Transition Perception/Recognition, Line Flow, Semi-structured Ops: validated in pilots, but replication conditions still maturing; exception handling & system modification costs remain high.

Long-term Complex Assembly, Exception Handling, General Multi-task: represent the capability ceiling, still in pilot/demo phase. Figure AI, Apptronik enter real production tests.

Short-term (closed-loop): Prioritize tasks with clear boundaries, quantifiable value, and manageable exceptions. Focus on real delivery, continuous operation, and customer acceptance.

Mid-term (reuse): Turn project-based solutions into task templates, delivery standards, and maintenance systems. Cross-site replication within the same task cluster is key.

Long-term (operations): Evolve from selling hardware/projects to continuous operation, data feedback, version iteration, and cross-site scaling. Revenue structure shifts toward platform and operational services.

FanTuo: Physical AI data & software, providing full-chain data collection-annotation-FunSim simulation-training ground, helping power grid clients build embodied AI labs, reducing personnel exposure risk.

Zongwei: Smart magnetic drive conveyor system, upgrading line-flow to a programmable, schedulable platform, serving AI Infra upstream hardware manufacturing, improving cycle time and flexible changeover.

The core contradiction in Industrial Embodied AI: supply-side emphasizes general intelligence and frontier forms, while demand-side needs stable, verifiable, and replicable task-closed-loops. The industry is moving from "capability showcase" to "task delivery", with value shifting to system integration, data loops, and continuous operations. The first successful scenarios are not the most complex but those with clear boundaries, manageable exceptions, and quantifiable ROI — logistics and fixed-station tasks. In the long run, data loops and platform capabilities will determine whether a company can move beyond single projects to replicable task systems.

The industry chain is repositioning; midstream integrators evolve from engineering executors to task-operation partners, and software infrastructure (simulation, orchestration, O&M platforms) becomes a new value control point.