Early-stage Project | A Former Horizon Product Head is Focused on "Pick-and-Place" Actions, Aiming for 100 Units of Wheeled Robots to Be Shipped This Year

Author: Ou Xue

Editor: Yuan Silai

At a time when the embodied intelligence industry is generally obsessed with bipedal humanoids and simulation training, a company has chosen a different path: focusing on the warehousing and logistics scenario, and tackling the "pick-and-place" actions that account for 60% of labor costs with a "wheeled chassis + dual arms". This is an embodied intelligence enterprise incubated by the Nanjing Institute of Software Technology of the Chinese Academy of Sciences - Zhizai Future.

Zhizai Future was founded in Nanjing in November 2025. Its founder, Sun Junkai, once served as the general manager of the intelligent cockpit product line at Horizon, promoting the mass production of millions of terminals and having product design and mass production experience from scratch. In its early days, the company was incubated under the Chinese Academy of Sciences system as an "embodied intelligence research group" for two years and was independently registered at the end of 2025.

The core of the generalization problem of embodied intelligence in real environments lies in the Sim2Real gap. Traditional offline reinforcement learning relies on simulation data and has a low deployment success rate; online reinforcement learning has high accuracy but a long learning cycle, making it difficult to implement in e-commerce warehouses with millions of SKUs.

Zhizai Future innovatively introduced the Human-in-the-Loop online reinforcement learning method, deeply integrating the immediate correction ability of humans with a unified reinforcement learning goal and opening up the key path from imitation learning to autonomous exploration. Based on this method, with only a small amount of demonstration data and a short period of online learning, the task success rate can be significantly improved, achieving an order-of-magnitude improvement in sample efficiency compared with traditional paradigms.

The company's first-generation intelligent robot, Armstrong, has been verified in a leading domestic logistics enterprise. The second-generation model, Armstrong Pro, was launched in the first half of 2026 and has successfully entered the warehouses of a Fortune 500 foreign pharmaceutical company for operation. In 2026, the company targeted to ship 100 units, which is estimated to account for nearly 40% of the market share in the industry.

Sun Junkai told Yingke that Zhizai Future's robots can "quickly enter the warehouse without modifying the warehouse and can be used for multiple purposes", completing operations such as shelving, picking, and inventory taking with "zero transformation cost" in the warehouse. The return on investment cycle for customers is about 2 - 3 years.

In the next 3 - 5 years, Zhizai Future has a clear roadmap. Sun Junkai revealed that from 2026 to 2028, the company will deeply cultivate the warehousing and logistics field and iterate the infrastructure model for logistics scenarios; in the medium and long term, it will reduce the generalization ability accumulated in the B2B market to the retail and home service sectors.

Sun Junkai explained: "The goods in warehouses - clothing, food, cosmetics - can also be found in supermarkets and homes. The ability to pick up packages can be almost directly transferred to the home organization scenario. We believe that robot butlers can pick up and unpack packages at home and organize items well. Therefore, we think that warehousing and logistics is the necessary path to the home."

The following is an excerpt from the dialogue between Yingke and Sun Junkai:

Yingke: How strong is the 'pick-and-place' demand in the warehousing and logistics scenario? Why not use humans?

Sun Junkai: A leading logistics enterprise has officially announced that it will achieve a fully unmanned warehouse within 8 years. The demand is much stronger than we originally thought.

The last mile of warehousing, that is, taking items from the storage bins and putting them into the order boxes, accounts for more than 60% of the labor cost, and there are often hundreds of thousands or even millions of SKUs. Traditional automation simply cannot achieve absolute generalization.

Large models are exactly good at generalization, and this is a scenario where the technology can be put to good use. The annual cost of a human worker is 50,000 - 100,000 yuan, while a robot can recoup its cost in just 2 - 3 years. The rigid demand of leading enterprises is already very clear, and the demand in the sinking market will gradually be released as the cost decreases.

Yingke: How to start the data flywheel in a real warehouse with the minimum amount of data outside the simulation environment?

Sun Junkai: The key lies in the consistency strategy. We deeply integrate the immediate correction ability of humans with a unified reinforcement learning goal and only need to collect a small amount of data and make fine-tuning for complex scenarios. In this way, the effective utilization rate of data is the highest, and the maximum generalization can be achieved with the least amount of data.

Yingke: Why use wheeled robots instead of bipedal ones?

Sun Junkai: First of all, the implementation in the B2B market ultimately follows the logic of cost - accounting - how many people are replaced, what is the labor - efficiency ratio, and how long is the return on investment cycle. Bipedal robots have a more complex mechanical structure and more degrees of freedom, and the system stability decreases exponentially; moreover, the total domestic shipment of bipedal robots is currently less than 10,000 units, and the supply chain cannot reduce costs.

Second, the probability of system failure for a robot with 20 degrees of freedom in the whole body and one with 60 degrees of freedom is not on the same level. B2B customers have high requirements for accuracy, efficiency, generalization, and reliability. At this stage, using bipedal robots is actually "using a sledgehammer to crack a nut" and has a high failure rate.

Third, the cerebellar motion control complexity of bipedal robots is much higher than that of wheel - arm robots. The industry has basically converged to the wheeled configuration that can be lifted or folded.

Of course, I believe that the ultimate form will definitely be bipedal robots with two dexterous hands. Because the most general configuration in the physical world must be like a human, capable of adapting to various environments and tools. However, this requires companies like Tesla and Unitree to push the production volume to the million - unit level and reduce the cost of bipedal robots to tens of thousands of yuan. By then, the supply - chain advantage of wheel - arm robots will not be obvious. But for now, it's too early.