The intra-logistics robotics industry has developed for more than 30 years, and even with the cost reduction of core components such as 3D radars, the market penetration rate remains at a low level. In 2025, the sales volume of industrial forklifts reached 1.45 million units, while the number of unmanned forklifts was only about 32,000 units with a penetration rate of less than 3%. Facing the industry phenomenon of price competition and parameter involution, Tao Yikun, Chief Scientist of Hangcha Group with 15 years of experience in the mobile robotics industry, has an in-depth judgment on the industry’s development logic and points out the core direction of industrial breakthrough with the combination of embodied intelligence and intelligent driving technology.
Tao Yikun believes that the industry’s current development predicament is not due to the exhaustion of market increment space, but is restricted by three major structural factors in the industrialization process: high deployment cost, poor scene adaptability and over-reliance on structured instructions. The landing of mobile robots in industrial scenarios is highly similar to the early stage of autonomous driving that relied on high-precision maps, requiring the construction of on-site infrastructure and resulting in high costs that exceed the bearing capacity of ordinary factories; standardized products are faced with highly fragmented application scenarios, and a large number of customized codes lead to “rule explosion” which makes the system difficult to maintain and reuse; natural semantic instructions can only be executed after being parsed into structured coordinate instructions by upper systems such as WMS, and the robot lacks independent perception and decision-making capabilities in complex environments. The superposition of these factors has become the key reason for the slow improvement of the industry’s penetration rate, and the fundamental solution lies in the bottom-level technological and industrial paradigm transformation.
Against this background, embodied intelligence has become the core starting point for the industry to realize paradigm transformation, and the autonomous driving industry that has completed large-scale commercial landing has provided a mature reference path for this transformation. Tao Yikun points out that the core breakthrough of autonomous driving in the past decade is not only the upgrading of a single point technology, but the construction of a complete “carrier-data-model” positive cycle, which has realized the leap from rule-driven to data-driven. After four stages of development, autonomous driving has formed a development model in which hardware carriers continuously obtain real-scene data, train models through massive data, and then feed back to upgrade system capabilities, and this model is also applicable to the development of the intra-logistics robotics industry. At present, mobile robots are still in the early stage of the autonomous driving evolution path, and the shift from AMR to EIMR (Embodied Intelligent Mobile Robot) driven by embodied intelligence is to make the industry move from “rule system” to “data driven” and from “scene adaptation” to “capability generalization”. On the one hand, embodied intelligence endows mobile robots with stronger perception, understanding and decision-making capabilities to adapt to open and complex industrial environments; on the other hand, industrial mobile robots provide continuous real-scene data for the iteration of embodied intelligence as a carrier with large-scale landing potential, forming a mutually promoting positive cycle.
Tao Yikun further clarifies the specific path of autonomous driving technology migration to the intra-logistics robotics field, and divides the opportunities and challenges in the migration process into “dessert” and “hard dish” with vivid metaphors. The most feasible and deterministic “dessert” is the “perception-heavy, map-light” technical paradigm. The current mobile robot deployment relies on SLAM to build maps and teach operation points with precise 3D coordinates, and the high cost of high-precision map production and maintenance has become a major constraint on large-scale landing. The autonomous driving industry has explored a path of replacing pre-drawn high-precision maps with real-time generation of local environment information through online perception networks such as BEV/OCC in urban scenarios, and the migration of this paradigm can greatly reduce the deployment cost of mobile robots and break the shackles of static maps, which is the most immediate breakthrough direction for the industry at this stage. Meanwhile, the industry also needs to face two core “hard dishes” unique to industrial scenarios: one is the magnitude gap in precision and granularity between the autonomous driving technology paradigm and industrial embodied perception. Autonomous driving in road scenarios only requires sub-decimeter level precision, while industrial operations such as forklift fork insertion and warehouse location alignment require sub-centimeter level control precision, and this gap cannot be bridged by simple technical transplantation and needs continuous engineering breakthroughs; the other is the data efficiency problem brought by the high customization and small-batch attributes of the industrial robot industry. Unlike the autonomous driving industry with unified product configuration and massive sales volume, the industrial robot industry has great differences in body, sensors and operation processes in different scenarios, resulting in high data collection cost and difficult cross-platform reuse. Constructing cross-device platform data representation has become a core key to solve this problem, and this research needs to be upgraded to an industry-level strategic issue.
The transformation of technical paradigm will ultimately lead to the reconstruction of product form, and the essential difference between AMR and EIMR lies in the fundamental change of underlying logic rather than superficial functional upgrading. Tao Yikun points out that AMR is rule-driven, relies on high-precision maps and executes coordinate instructions, and is essentially an automated system that cannot be separated from upper systems such as WMS and RCS; while EIMR is data-driven or hybrid-driven, adopts the “perception-heavy, map-light” route, runs on semantic instructions, and takes replacing manual labor as the core goal to realize the “single machine autonomy” of X=ZERO in the V2X context. Based on this, the evolution direction of the industry is from group (system) intelligence to single machine (distributed) intelligence. The multi-machine scheduling that has been highly recognized in the industry is essentially an assignment problem of operational research with a limited technical upper limit, and the real core competitiveness of the industry is to realize the ability of robots to complete complex operations independently without networking and relying on external systems, which is also the key to large-scale replication of products.
For the industry to realize the paradigm leap from AMR to EIMR, Tao Yikun proposes a pragmatic transition path of hybrid intelligence. Pure rule-driven can realize rapid landing of products with low threshold, but has a very low development upper limit; pure data-driven has unlimited growth space, but requires huge early data costs and the short-term effect is not obvious. Hybrid intelligence combines the advantages of both, relying on the rule-driven system to quickly land products and obtain real-scene operation data, and at the same time retain the long-term growth potential of data-driven intelligence. He emphasizes that data engineering is not only an internal matter of the algorithm team, but must be upgraded to the top-level strategy of the enterprise, because only when the product runs in the real scene can the data flywheel be leveraged to promote the continuous iteration of the system.
Referring to the autonomous driving classification system, Tao Yikun also draws a clear evolution roadmap for EIMR: the current “L4-level unmannedization” of AMR relying on V2X is essentially a “high-cost pseudo-L4” realized by stacking external systems. EIMR needs to first break the external dependence on V2X and return to L2 level to realize reliable human-machine collaboration; then climb to L3 level with the iteration of embodied intelligence to realize fully unmanned operation in conventional scenarios with only sparse manual intervention in special cases; and finally move towards the real L4-level intelligence with strong environmental adaptability, which can understand and parse natural language instructions and realize full-process unmanned operation in non-standard scenarios such as dense and narrow spaces.
Tao Yikun’s judgment on the industry is calm and pragmatic. He believes that the path of combining embodied intelligence with mobile robots to realize the paradigm shift from AMR to EIMR has a time window, technical feasibility and industrial foundation support, and is not a blind exploration of the future. When embodied intelligence is deeply integrated with the product paradigm of intra-logistics robotics, the industry will break through the current development predicament and enter a new stage of large-scale development.
