The year 2026 is being defined by the industry as the "delivery year" for embodied intelligence. While many companies remain trapped in the Sim-to-Real gap—struggling with insufficient real-world robot stability, a scarcity of high-quality physical interaction data, and weak cross-scenario generalization capabilities—the competitive benchmark has shifted from exhibition hall demonstrations to production line deployment. Leveraging over sixty years of industrial experience and a full-stack, self-developed system, Ningbo Pia Automation Holding Corp. (SH.688306) has built a complete closed-loop process from "single-unit capability testing to multi-level scenario validation to production-line scale deployment," utilizing the scenario resources of its global network of 100 factories. This has established a validation landscape covering high-end manufacturing, retail, and general services.
Starting with Single-Unit Pre-Job Testing: Solidifying the Perception-Action Loop Foundation
The first hurdle for embodied intelligence deployment is closing the "perception-decision-execution" loop within a real physical environment. Ningbo Pia Automation has established a standardized real-robot testing system to validate two core capabilities: precision force control and teaching-learning. In the retractor housing feeding scenario at the group's JSS factory, the joint R&D lab with Zhiyuan achieved a flexible assembly success rate of 99% and a fastest cycle time of 12.97 seconds. In a steering mechanism assembly task, the company partnered with a global electric vehicle giant on a concept verification project, overcoming challenges in whole-body multi-degree-of-freedom hybrid force-position control and integrating tactile dexterous hands to achieve multi-specification screw grasping and multi-angle precision peg-in-hole insertion. For the retractor housing assembly, self-developed dual-arm admittance force control and real-robot reinforcement learning algorithms solved the challenge of synchronizing three pins, achieving a stable assembly success rate of 99%. A flexible circuit board insertion scenario validated a transferable skill learning paradigm: by collecting demonstration data through isomorphic master-slave teleoperation and iterating with reinforcement learning, the robot autonomously mastered the insertion skill, converting manual process expertise into a reusable capability.
Multi-Level Scenario Validation: From Workstation Closed-Loop to Multi-Robot Collaborative Lines
An automated vibratory bowl feeder project achieved a complete unmanned cycle operation. Relying on the self-developed PIAVision follow-up vision and autonomous path planning, the system autonomously completed the entire process of bin picking, transfer, dumping, and resetting, validating the reliability of 7x24 continuous production. At WAIC, the world's first multi-robot collaborative production line for automotive BMS precision assembly was showcased. System-level validation results were presented at the 2026 WAIC—the world's first BMS controller pilot production line featuring six self-developed G2 series embodied robots working in tandem. This line covers complete processes including raw material transfer, precision assembly, flexible labeling, bin storage, intelligent inspection, and predictive maintenance, all orchestrated by the self-developed Jarvis Industrial Kit (JIK 2.0) in conjunction with PLC global scheduling. Tens of thousands of hours of real-robot testing show the line's comprehensive stability rate reaches 99.9%, with single-process cycle time controlled within 70 seconds. Dr. He Chuan, head of the research institute, stated that the production line has completed a full process cycle and is ready for rapid replication in 3C precision assembly, medical devices, and smart warehousing. The layered validation across five key industrial scenarios has charted a complete path from single-point breakthroughs to full production line scale-up.
Data Infrastructure: Building the Industrial Embodied Intelligence Data Flywheel
Global real-world physical interaction data amounts to only about 500,000 hours, leaving a massive order-of-magnitude gap compared to the millions of hours required for training. Ningbo Pia Automation has developed the Primus Forge data platform and the Primus Ego head-mounted collection device, constructing a complete "collection—quality check—annotation—training—evaluation—deployment" pipeline. On the data collection front, the Primus Ego head-mounted device uses a 270° first-person perspective to efficiently capture workshop process data, directly converting manual process expertise into trainable data assets. This device is adapted for the complex conditions of industrial manufacturing and can synchronously capture multi-modal data including visual, force, and motion trajectory information, significantly improving the efficiency and quality of real-world physical interaction data acquisition. On the data processing side, the Primus Forge platform employs a VLM large model to replace 80% of manual annotation, greatly enhancing processing efficiency and annotation consistency, and enabling efficient conversion from raw data to standardized model training data. In terms of the data ecosystem, the company, in collaboration with Bodon Intelligence and SJTU's MINT Lab, has open-sourced the RW-RL-Dataset dataset, initially containing over 1,000 hours, with plans to expand it to 3,000 hours by the end of 2026, filling a gap in real-world interaction data for domestic industrial robots.
Extending Scenarios: Industrial Tech Prowess Radiates into Retail and General Services
Leveraging the high-precision perception, force control, and task planning capabilities honed in industrial settings, Ningbo Pia Automation is extending its validation landscape into commercial sectors. In retail scenarios, in partnership with retail technology partners, it has accumulated over 2,000 hours of shelf operation data, covering atomic actions like picking, restocking, and flexible arrangement. In the general services domain, the company is collaborating with ecosystem partners to build a long-sequence task human-robot interaction dataset, aligning natural language instructions with robotic action logic to optimize the VLA model's ability to understand complex tasks.