Shanghai Sazhi Intelligent Technology has closed a Pre-B round of several hundred million yuan, bringing in three investors: Jiantou Investment Co., Ltd., Nantong Investment Management and Yunqi Partners. The money is earmarked for four things: large-model R&D, iteration of the multi-arm robot platform itself, production line expansion, and a deeper push into global markets.

The shape of the company's product is worth spelling out first. From single-arm to dual-arm, four-arm and six-arm, everything it builds is a multi-arm robot, meaning several robotic arms mounted on one mobile body. That choice traces back to a very concrete constraint on the factory floor.

One Robot, Four Production Lines

Industrial manufacturing is moving toward greater flexibility and complexity. Conventional industrial robots are fixed in place and single-purpose: program one arm and it does one job. General-purpose embodied robots have drawn plenty of attention over the past two years, yet they often fall short of industrial-grade precision and determinism.

Sazhi is going after the gap in between: precision assembly, dexterous packaging, long-sequence testing, the kinds of steps that demand flexibility and accuracy at the same time.

So why one multi-arm robot instead of several single- or dual-arm machines working side by side? Zhang Jianzheng's answer comes from customer sites. Space on existing production lines is often extremely tight, yet output has to stay high. Deploying multiple two-arm robots eats floor space and makes tight coordination very hard.

The six-arm answer is what the company calls spatial heterogeneity: two arms each at the front waist, rear waist and shoulders, with every arm getting its own independent three-dimensional working envelope. That lets the robot cover several nearby production lines through a full 360 degrees, running two workstations in parallel and alternating four arms.

Zhang offered a concrete picture of the work. When the three arms on the left side operate together, the lower two each hold a different part for assembly while the upper one wields an electric screwdriver or glue gun for fastening or dispensing. In packaging, bagging, inserting the manual, sealing, boxing and capping all require several hands working at once; six arms can finish the sequence in one pass.

Keeping Six Arms Out of Each Other's Way

More arms bring more trouble. Compared with four arms, six raise the degrees of freedom by 60%. Degrees of freedom can be read as how many poses an arm can strike: the more joints, the more corners it can reach, but also the more avoidance relationships that have to be computed.

In cramped spaces, those extra degrees of freedom have to solve three problems: dynamic isolation of safety zones, arm-to-arm collision avoidance, and motion coordination under complex processes. All three place steep demands on computing power and on how well the algorithms actually work.

By Zhang's account, this spatially heterogeneous six-arm configuration is unique worldwide. Task complexity rises exponentially compared with four arms, and efficiency improves by 50%.

The four-arm machine takes a different route: a spatially symmetrical layout, with one brain controlling four arms and coordination algorithms doubling the efficiency of parallel multi-arm operation versus the previous generation.

On whether more arms are always better, Zhang has an analogy: more is not automatically better, much as a table needs at least three legs to stand, four or five are fine, and beyond that it stops meaning anything. Grounded in spatial configuration theory, he sees six arms as the most sensible layout available today.

Imitation Is Not the Same as Understanding Cause and Effect

Multiple arms solve the body. Next comes the brain. The industry currently leans on VLA models, or vision-language-action models: see the scene, parse the instruction, output the motion directly. Sazhi's read is that this path has limits. End-to-end reflexive imitation lacks any grasp of physical causality, so generalization breaks down as soon as the scene changes, and the process cannot be explained.

Reflexive imitation is easy enough to picture. After seeing enough demonstrations, the model copies what it saw whenever it encounters a similar image, but it does not know why it is doing so, nor whether the behavior still holds with a different part or a different angle.

Sazhi's answer is its in-house industrial world model, Xuanshu · SageNexus, unveiled this year at WAIC. The company says the core breakthrough is internalizing physical causality, shifting the system from passive response to active reasoning.

The architecture works in two stages. An encoder maps data of entirely different natures, including vision, point clouds, force feedback and process documents, into a single physical-semantic space. A decoder then translates high-level task intent into executable motion sequences, running physical feasibility checks during translation.

That check is the crucial part. Once a motion is generated, the system first judges whether it is physically achievable, whether it will cause a collision, and whether it exceeds process constraints; if it fails, the motion is recomputed. Large-model hallucinations get stopped here before they turn into an unsafe command.

The foundation is data. Built on 360TB of multimodal real-world industrial data and three-stage training, the model has accumulated a skill library covering more than 2,000 scenario skills, and meets the determinism requirements of industrial settings with reliability above 99.9%. Zhang sums the whole setup up as a closed loop of five elements: data, hardware platform, control, model and scenario.

A Path That Does Not Depend on Stacking Chips

Embodied intelligence has a common affliction right now: bring in a large model and computing power costs climb with it. Zhang's view is that piling on more computing resources only drives up what customers have to spend, and that raising algorithmic efficiency to reduce dependence on brute-force computation is what genuinely works in deployment.

There are three layers to how they do it. The first is data, drawn entirely from actual industrial settings; clean data means less wasted effort in training. The second is optimization of the algorithm architecture. The third is deployment: once the model is trained, some skills can be pushed to run directly on the device at the edge rather than returning to the cloud every time.

Supporting that deployment approach is a three-loop architecture: a slow loop in the cloud for deep thinking, a fast loop at the edge for quick reaction, and a safety-floor loop as backstop. A closed loop here means acting, checking the result, then adjusting the next step; the difference between slow and fast is that planning can take its time to be thorough, while real-time correction has to be quick.

The reason for splitting it into three loops is practical. On long, complex tasks, relying entirely on the cloud is too slow, and relying entirely on the edge leaves too little compute.

The First Half Already Matched All of Last Year

The commercial numbers say more than the technical description. Sazhi now covers four major industries, 3C semiconductors, automotive, energy, and food and pharmaceuticals, with more than 50 customers among Global 500 or publicly listed companies.

Zhang said the company had already completed its full-year 2024 target in the first half of this year, with business for the year roughly doubling at a growth rate between 100% and 120%.

Overseas has become the second growth curve. Last year much of the company's business was in Mexico; this year it has begun a full push into Southeast Asia, with work underway in Vietnam, Indonesia, Malaysia, Thailand and Singapore, and large orders gradually landing in Europe as well. Overseas letters of intent reportedly exceed 500 million yuan.

To absorb that demand, the company's plant in Shanghai's Maqiao AI pilot zone is running at full capacity, and it plans to expand to roughly four times its current footprint within the next three years.

Next Up: Hundreds of Robots Working as One

Sazhi divides its roadmap into three stages: from individual intelligence to swarm intelligence to ecosystem intelligence, corresponding to versions 1.0 through 3.0.

Zhang said the company has already entered the swarm intelligence stage, with early swarm control algorithms allowing several hundred robots to work jointly inside one factory. The next step is to keep iterating on swarm intelligence and strengthen multi-agent coordination, aiming for tighter cooperation and more precise control.

Ecosystem intelligence comes after that. At that point the task is not just managing hundreds or thousands of robots, but bringing the factory's other hundreds or thousands of machines under unified control as well.

The investors' reasoning points the same way. Nantong Investment Management cited the company's accumulated work in high-quality industrial data, large models aimed at the physical world, and encoding-decoding systems for the factory floor. Wang Feng, investment director at Jiantou Investment, argued that the scarcity of genuinely effective industrial deployments and the data that comes with them is the core bottleneck facing embodied intelligence companies today. Yunqi Partners pointed to two hard constraints on the factory floor: limited production line space paired with complex processes, and stability over long periods of operation.

From one robot tending four production lines to hundreds of robots coordinating inside a single factory, the hard part of this path was never the demo. What industrial customers pay for is whether the same set of motions holds steady after a change of part number, workstation or shift.