机器人“实习”几天就能上岗，还会越干越熟？这家清华系创企只用了..

Robots can start working after just a few days of “internship,” and they get better with experience? This Tsinghua-affiliated startup only used..

Robots must not only learn to complete tasks, but also learn to do them better.

Author | Wang Bo, Zhang Lin

At the 2026 China International Industrial Expo (referred to as the “Industrial Expo”), a red-and-black single-arm robot faced a pile of boxes of varying sizes and orientations, repeatedly performing mixed‑category box palletizing tasks.

However, upon closer inspection, “Jiazi Light‑Year” noticed that this live demo differed somewhat from other similar demonstrations. The robot does not follow a fixed trajectory to sequentially pick up boxes: during assessment, the arm reaches toward a box in front, then changes direction to take a higher box; during grasping, it adjusts the suction cup’s pose relative to the box; and when placing, it navigates around already stacked items.

Generally speaking, exhibition settings impose certain constraints on robot operation, such as the size of the exhibit area, lighting, and interference from surrounding people and equipment. Some demos that can be completed in a lab may encounter issues on site. However, the team that completed this demonstration used only about three days to collect data, two days for reinforcement learning, and the remaining day for inspection and adjustments.

Multi‑category box mixing palletizing operation

The single‑arm robot unveiled at the Industrial Expo is called Sculptor, and the embodied intelligence company behind it is Shenyi Robotics. Shenyi Robotics defines Sculptor as “the world’s first end‑to‑end model‑driven intelligent loading and unloading robot.”

Shenyi Robotics is a Tsinghua‑affiliated startup that has been founded for less than nine months. Founder and CEO Liu Kai, co‑founder and large‑model lead Luo Yu, and co‑founder and embodied‑system lead Huang Kangyao all graduated from the Computer Science Department of Tsinghua University.

In addition, Shenyi Robotics also announced for the first time its self‑developed embodied foundation model SEEN·E SE‑1 and the dual‑arm robot Edge for precision industrial manipulation.

Edge on-site demonstrated two tasks: sub‑0.1 mm precision insertion and interactive candy dispensing, which correspond to two key capabilities of embodied intelligence: fine physical manipulation and the understanding and execution of complex tasks, showcasing more possibilities for embodied intelligence in flexible industrial operations.

Sub‑0.1 mm Precision Insertion

Interactive candy packaging (randomly assign specific candy types for the robot to pack)

These three demos feature different robot forms and tasks, but they all point to the same question:Can robots shorten the “probation period” when taking on new tasks and continue to accumulate experience in real work?

At this year’s WAIC and WRC, the embodied intelligence industry showcased numerous grabbing, handling, and plugging demos. At the more production‑site‑focused Industrial Expo, customers asked concrete questions: Can it still handle a new batch of goods? How long does it take to commission in a new factory? Is continuous operation stable? When can the investment be recouped?

The Shenyi team completed a new mixed‑palletizing task in about a week, answering only the first question. What they really want to prove is whether this learning speed can be replicated across more customer sites.

1. One‑Week “On‑Job”, What Sculptor Learned

The on‑site demonstration of multi‑category box mixed palletizing by Sculptor may appear simple. Indeed, if the positions of the boxes are all fixed, engineers can pre‑program the industrial robot arm’s trajectory from point A to point B. If the box dimensions are identical, the stacking pattern is regular, and the operation cycle remains unchanged for a long period, traditional automation solutions can also deliver higher efficiency and stability.

But reality is not like that.

Liu Kai co‑founded the intelligent logistics robot company Jizhi Jia, experiencing a robot company’s journey from product definition and large‑scale delivery to going public. Before his next venture, he visited numerous logistics and industrial sites both domestically and abroad.

One scene that left a deep impression on him was a logistics hub in Hong Kong. In summer, the outdoor temperature exceeded 35 °C, and the interior of containers was even hotter, with workers handling boxes of varying weights shirt‑sleeve‑less. The cargo inside the containers was extremely chaotic, with no predetermined stacking plan, so workers had to identify, sort, and re‑stack items during handling.

Real scenes of different logistics transit centers

These tasks that have not yet been covered by automation in reality often share several characteristics: the object’s position constantly changes, the task involves some judgment, and the process is sufficiently labor‑intensive.

Existing robot handling solutions for container scenarios either involve robots that are too large to maneuver in confined spaces, or the robots merely transport goods mechanically without making effective judgments.

Sculptor has a working radius of 1.8 m, a maximum load of 35 kg, positioning or operation accuracy <2 cm, can accommodate boxes ranging from 10 cm to 80 cm, can operate continuously for up to 8 hours, and achieves a cycle time of 400–800 boxes per hour. It also features omnidirectional movement and some climbing and obstacle‑crossing capability.

This makes it especially suitable for cargo handling in container scenarios, and the Deep Insight team has already conducted some training and testing. In the demonstration at the Industrial Expo, although the scenario changed, Sculptor quickly adapted.

Luo Yu recapped the changes to Sculptor over the past week for “Jiazi Light‑Year”.

In multi‑category box mixing palletizing tasks, the robot must first decide the picking order. Luo Yu showed a before‑and‑after comparison of reinforcement learning: before training, Sculptor knows it needs to move boxes but may try to pull a box that is pressed down by a box above it, causing surrounding boxes to be squeezed or even the stack to collapse; after reinforcement learning, it sometimes changes its choice after extending its arm, opting to remove a higher‑positioned box first.

The second change appears in the grasp pose. Sculptor’s end‑effector suction cup is equipped with a camera; if the long side is oriented toward the short side during grasping, the camera can easily be occluded by the box, and the stacking structure may be incorrect. After training, it proactively adjusts the suction cup pose to align long side with long side.

The third change occurs during placement. Previously, the robotic arm occasionally brushed over other boxes from above, even pressing down on already stacked boxes; after reinforcement learning, it begins to route around from the side, finds a more reasonable path, and then places the boxinto the target position.

Reinforcement Learning Drives Self-Evolution

These actions are not pre‑written fixed trajectories for each box. The cardboard boxes on site are placed randomly, and the initial positions may differ each time; the robot obtains environmental information through a camera video stream, and the model generates subsequent actions. It is not just a single transport from point A to point B, but a long‑range task that includes identification, selection, grasping, collision avoidance, and stacking.

So, why can reinforcement learning change Sculptor’s way of picking, grasping, and navigating around boxes?

2. Get the robot quickly on the job,

also let it learn from the work

Shortening a robot’s “internship period” requires solving two interrelated problems:It must first understand the task and the physical world, and then adjust its behavior based on real feedback.

Deep Insight summarizes these two stages as“understanding” and “alignment”.

Core Mechanism of Self‑Evolution

“Understanding” is mainly achieved by the embodied foundational model SEEN·E SE-1 released by Shenyi at this Industrial Expo. SE-1 follows a layer‑by‑layer refinement approach: from semantic tasks to multi‑dimensional physical prediction, then generating continuous actions.

SEEN·E SE-1: Task Understanding, World Inference, and Smooth Operation

When faced with the instruction “place these cardboard boxes onto the pallet,” the model must first combine the previously completed steps and break the abstract goal into a series of consecutive sub‑tasks. Within the same sub‑task, multiple world‑model experts each predict future latent representations, depth, and key‑point trajectories, and the action module generates the robot’s continuous motions based on this information.

The SE‑1 model focuses on the physical changes directly related to robot actions. It needs to determine whether extracting a particular box would cause the stack to collapse, whether the robotic arm moving along its current path would result in a collision, and what orientation the box should be placed in.

After understanding the task, the robot must also align its behavior with the work objectives defined by humans.

Shenyi calls this process “alignment”: through real interaction and reward‑penalty feedback, reinforcement learning gradually eliminates failed actions and retains more effective strategies. The “value” here is very concrete, including higher success rates, fewer collisions, more accurate plug‑in, and more stable operation cadence.

The dual‑arm harness‑plugging demo provides an intuitive example. The positions of the harness and the connector are not fixed, and the system does not obtain XYZ coordinates in advance. When learning only from demonstration data, the robot knows that “it should insert the wire,” but after a misalignment it sometimes continues to push forward; after adding reinforcement learning, it will back off, realign, and try again when it detects a failure.

“Reinforcement learning is the key to improving a robot’s success rate during the insertion process,” Luo Yu told “Jiazi Light‑Year”, “We measured that after a day of reinforcement learning the success rate can rise from 60% to 90%, and in the past two days of on‑site testing at the exhibition, the success rate exceeded 96%.”

In real production sites, the amount of new data that can be generated is usually limited, so Shenyi introduced OMPO. When the environment, target location, or robot state changes, OMPO matches and calibrates the distribution of historical data with a small amount of on‑site data, allowing previously accumulated experience to continue to be used for value learning and policy updates.

OMPO

Asynchronous reinforcement learning is responsible for converting this experience into the next version of the policy. Robots generate new data on site, and the training system simultaneously draws on historical tasks and data from other robots to perform value assessment and update action policies, then deploys the new policy for the next round of interaction. The training process is decoupled from on‑site operation, allowing model updates during robot downtime or charging periods.

Asynchronous Reinforcement Learning

PRISM and TimelyDAgger further improve the efficiency of training data. PRISM identifies higher‑precision action phases within a task, such as the final segment when a cable approaches a socket, and increases the supervision weight for these actions; TimelyDAgger determines when the robot truly needs expert intervention, allowing human demonstrations to focus more on capabilities the model has not yet mastered.

PRISM

TimelyDAgger

These methods all point to what Shenyi calls “self‑evolution”.

Luo Yu emphasized that self‑evolution is a goal, not a single algorithm.Data generated during robot operation still needs to be filtered, trained, safety‑validated, and version‑deployed; robots will not modify their actions on their own during production without inspection.

In summary, SE‑1 handles task and physical‑world understanding, reinforcement learning corrects behavior based on outcomes; OMPO, PRISM, and TimelyDAgger each improve the utilization efficiency of past experience, fine‑grained motion data, and expert demonstrations, while asynchronous reinforcement learning incorporates these experiences back into the policy.

“Jiazi Light‑Year” discovered that,Deep Yi maintains restraint toward the concept of a “world model” itself.

“We should still think from first principles about what an embodied foundational model should actually do,” Liu Kai told “Jiazi Light‑Year”. In his view, a robot’s world model needs to understand the physical laws directly relevant to action. SE‑1 uses a lightweight world‑model MoE architecture, focusing the modeling on information such as force, depth, and object trajectories, rather than generating a pixel‑level future video.

This architecture also incorporates the task decomposition, reasoning, and cognition capabilities of VLM, and introduces a MoT (Mixture‑of‑Transformers) architecture at the motion‑generation layer, providing a foundation for reinforcement learning and continual update strategies.

For Shen Jie, the world model ultimately must deliver a few concrete outcomes: whether the robot can understand tasks more quickly, generate actions more accurately, and continuously improve in real‑world operation.

3. Nine months to delivery, why is this team so fast

At the Industry Expo, which is closer to the industrythe topic of “implementation” naturally concerns the audience and customers.

Liu Kai summarizes the deployment scenarios of Shenyi robots into two directions:Mixed handling and precise operation. The former is closer to heavy‑duty “general worker” tasks, while the latter has a certain “skilled worker” attribute.

Therefore, Shenyi has designed two robot forms: Sculptor for heavy‑load logistics tasks, and the dual‑arm robot Edge for precise operations. The robot bodies are redesigned according to the tasks, while the underlying models and learning mechanisms are fully reused.

This approach can be summarized as“1 brain + 2 robot models + multiple scenarios”.

Shenyi “1+2+N” Strategy

From its founding to delivering a foundational model and two robots, Shenyi took less than nine months. Behind this speed are cutting‑edge model and reinforcement learning research on one side, and robot productization and large‑scale delivery experience on the other.

Liu Kai, from the Department of Computer Science at Tsinghua University, has over 15 years of experience in robotics technology and systems engineering. He spent ten years at Jizhi Jia working on logistics robots, covering product definition, hardware and system development, mass production delivery, and global commercialization. Luo Yu holds a Ph.D. in Computer Science from Tsinghua University and previously worked at Huawei Noah’s Ark Lab on foundational models, multimodal models, and reinforcement learning. He was selected for the “Huawei Genius Youth” and “Tencent Qingyun Program.” Huang Kangyao also earned his doctorate from Tsinghua’s Computer Science department and is responsible for robot systems, interactive reinforcement learning, and control.

In addition, Professor Liu Huaping from Tsinghua’s Department of Computer Science and a recipient of the National Outstanding Youth Fund serves as the chief scientific advisor for Shenyi Robotics. As one of the earliest domestic scholars researching embodied intelligence, his academic papers have been cited over 20,000 times. In 2025, he was also selected as an IEEE Fellow for 2026, the highest honor bestowed by IEEE.

This set of capability combinations determines the starting point of Deep Insight’s R&D:The model team looks for problems in real operational tasks, while the industry team must embed model capabilities into deliverable whole machines and systems.

Liu Kai summarizes the team’s capabilities as “from industry to model, then to reinforcement learning and systems,” and the team’s mindset is “lightly armed.”

“Jiazi Light‑Year” discovered that Deep Insight does not wait for models to become sufficiently “general” before seeking scenarios. The team first entered logistics and industrial sites, looking for processes that still rely on manual labor yet where customers can calculate a clear ROI, and then designed robot forms, models, and training pipelines around those tasks.

“We will not launch a product that isn’t ready for delivery,” Liu Kai told “Jiazi Light‑Year,” “the newly introduced Sculptor robot can be delivered to customers next month.”

Liu Kai disclosed that Shenyi currently has more than ten logistics and industrial POC (proof‑of‑concept) projects underway, covering scenarios such as loading and unloading, automotive parts, and 3C components.The next step is to see whether these projects can run stably and be replicated at more customer sites.

His past industry experience makes Liu Kai especially sensitive to the gap between “launch” and “delivery.” He clearly understands that if a company announces a product and it still cannot be delivered months later, customers’ expectations will quickly be eroded.

This is another example of Shenyi demonstrating pragmatism: they not only discuss what tasks the robot can perform, but also talk about deployment cycles, operation cadence, continuous runtime, and return on investment.

According to Liu Kai’s calculations, in logistics scenarios, Sculptor’s payback period is about 12 to 18 months.

Sculptor transporting goods

Liu Kai does not believe embodied robots will quickly replace industrial arms and collaborative robots. For tasks with fixed positions, stable processes, and clear cycles, traditional automation remains more mature and efficient. Embodied intelligence is better suited for work that involves multiple process connections, continuously changing environments, and requires some judgment ability.

This outlines a relatively clear commercial boundary for Deep Insight: they are looking for the gap between traditional automation and manual operation.

However, there are still multiple hurdles between preparing a trade show demo in a week and delivering stable implementations at multiple customer sites.

Out of a dozen‑plus POCs, how many will ultimately convert into large‑scale orders, how many engineers need to be投入 for on‑site adaptation, whether data from different customers can help the model continuously improve, and whether a 12‑ to 18‑month payback period can hold up in real operations—these questions will determine whether 深诣 can make the next leap.

As the industry’s focus shifts gradually from financing and demos to orders, delivery, and ROI, 深诣 still has to face the scaling challenges that any startup robotics company encounters.

“甲子光年” asks Liu Kai: “As a startup, what can be considered a stage‑level safety?”

Liu Kai replies: “Continuously supporting the company’s operations with real customer orders.” He especially emphasizes the word “real.”

As soon as the words fell, Sculptor took a cardboard box from the pile of goods, turned around, and steadily placed the box onto the tray.

(Cover image and in-article illustrations, video source: Shenyi Robotics)

END.