Sharpa has raised more than RMB 4.5 billion (USD 668 million) in a funding round that values the robotics startup at RMB 22 billion (USD 3.3 billion) post-money.
Its investors include Alibaba Group, Meituan, Tencent, JD.com, and Transsion, as well as investment firms HSG, Qiming Venture Partners, DragonBall Capital, and Luminous Ventures.
Sharpa said it will use the funding to accelerate R&D, recruit and develop talent, and move its general-purpose robots from technical validation toward commercial deployment.
Founded in 2024, Sharpa focuses on dexterous manipulation, or the ability of robots to use their hands to handle objects and perform complex physical tasks. It was co-founded by Hesai Technology CEO Li Yifan, CTO Xiang Shaoqing, chief scientist Sun Kai, and others.
The company is developing embodied intelligence systems that combine software, models, sensors, and robotic hardware. Its longer-term goal is to build general-purpose robots capable of operating in environments designed for people.
Sharpa said its robots are intended to take on repetitive, physically demanding, and intensive work rather than replace human workers altogether.
The company currently serves robot original equipment manufacturers (OEMs), research institutions, and food service customers. It eventually plans to expand into more commercial applications and, later, the home.
Its latest attempt to demonstrate that technology is taking place inside a Dairy Queen store in Shanghai.
On August 29, Sharpa and Dairy Queen, or DQ, jointly opened what they describe as the world’s first restaurant capable of year-round autonomous robotic operation without modifications to the existing store environment. The deployment is at DQ’s Wujiang Road outlet in Shanghai, which is operated under CFB Group.
Sharpa said the project marks an industry first in which a robot autonomously performs a complex production workflow in a live commercial environment.
The robot handles the entire 55-step process for fulfilling a “Blizzard Treat” order, from receiving the order and preparing the dessert to handing it to the customer. Sharpa says the system is designed to operate throughout the store’s daily business hours, from 10 a.m. to 10 p.m, without continuous human control.
According to Sharpa and DQ, the deployment also differs from many previous commercial robotics projects because the robot uses the same standardized equipment, ingredients, tools, and procedures already used by human employees.
The companies said the project is intended to address three recurring challenges in commercial robotics: dependence on customized equipment, difficulty performing complex tasks over long periods, and insufficient reliability for sustained commercial use.
For Sharpa, the more important test is therefore not whether a robot can complete one task once, but whether it can repeat a long sequence reliably under real operating conditions.
Li Yifan, a Sharpa co-founder, said commercializing robots requires them to perform complex work over long periods and across changing situations without requiring environments to be redesigned around them. Sharpa chose a restaurant as an early test of whether those capabilities can eventually transfer to other commercial settings.
Working in a store built for people
DQ’s Blizzard Treat provides a demanding test case because its production process relies on standardized equipment, ingredients, procedures, and quality requirements.
For Sharpa’s robot to operate within that system, the company says it had to address three areas: adapting to an existing human workspace, completing a complex workflow autonomously, and maintaining reliable operation over extended periods.
The first is what Sharpa calls “zero retrofit.”
At the Shanghai store, Sharpa said DQ’s equipment, ingredients, procedures, and operating standards remain unchanged. The robot uses tools designed for human employees and works within the store’s existing layout and workflow.
That approach is intended to make the technology less dependent on customized sites. If robots can use equipment already found in restaurants and other commercial locations, Sharpa argues, deployment could become easier to replicate.
Using existing equipment, however, addresses only part of the challenge.
The second test is whether the robot can complete a complicated production process autonomously.
Preparing and delivering a Blizzard Treat involves 55 interdependent steps. An error at one stage can affect later parts of the workflow.
Sharpa said its system links task planning, action execution, state assessment, and error correction so the robot can work through the sequence without continuous human intervention.
The third challenge is reliability over time.
During commercial operation, the robot must contend with changing order volumes, customer traffic, and other variations in the store environment. Sharpa says prolonged operation is intended to test whether the technology can produce consistent economic value rather than simply demonstrate technical capability.
Together, the three requirements form Sharpa’s proposed framework for commercial deployment: use existing environments, operate autonomously, and remain reliable over extended periods.
Breaking a 55-step process into smaller tasks
Sharpa has developed a technology stack spanning artificial intelligence models, sensing systems, and robotic hardware for dexterous manipulation.
At the DQ store, the system combines self-correction, tactile sensing, and a multimodal world model to complete the 55-step preparation workflow.
Long-horizon tasks present a particular problem for robots because small errors can accumulate as a sequence progresses.
Sharpa said that training the entire 55-step workflow as a single trajectory would make it difficult to account for every possible deviation. As the state of the environment changes, the number of possible paths through the task can expand rapidly.
Instead, the company breaks longer workflows into smaller subtasks that can be planned, checked, and repeated when necessary.
After each stage, the system evaluates the resulting state. If the outcome differs from what was expected, the robot can retry the action or perform an error-recovery procedure before moving on.
That creates a recurring loop of planning, execution, verification, and correction.
Tactile sensing addresses what Sharpa describes as the “last millimeter” of physical manipulation, where small differences in grip, friction, pressure, or object position can determine whether a task succeeds.
CraftNet, Sharpa’s internally developed hierarchical end-to-end model, incorporates tactile inputs. Combined with the multi-finger Sharpa Wave robotic hand, the system adjusts finger position and applied force in response to tactile and force feedback.
Sharpa said tactile sensing is involved in 98% of the 55 steps.
Pulling out a paper cup, for example, requires the robot to respond to friction and resistance. During high-speed mixing, it must adjust its grip as vibration and force change. Performing DQ’s signature upside-down flip requires the system to adapt as the cup’s orientation and center of gravity shift.
Sharpa’s multimodal world model is intended to determine how the robot should proceed as those conditions change.
The system combines vision, language, tactile information, force feedback, and internal robot-state data such as joint angles and proprioception, or the robot’s awareness of the position and movement of its own components.
According to Sharpa, the model uses that information, together with the system’s previous observations and actions, to predict the likely result of an action and determine what to do next.
In principle, the different components address separate parts of the same problem. Self-correction is intended to prevent errors from accumulating over a long task. Tactile feedback helps control physical contact, while the world model helps the robot assess its state and choose subsequent actions.
Together, they create a loop of perception, prediction, execution, verification, and correction.
The DQ deployment is intended to test whether that system can perform one relatively complex workflow reliably enough for commercial use.
From one workflow to broader applications
Sharpa’s longer-term ambition is to transfer the same capabilities across different commercial settings.
A key part of that approach is designing its robots to use objects and equipment already made for humans.
The size, form, and degrees of freedom of the Sharpa Wave dexterous hand are modeled on the human hand. That allows it to manipulate objects such as cups, cabinet doors, and spoons without requiring specialized tools.
This hardware design, Sharpa said, could make individual manipulation skills easier to reuse across different products, commercial kitchens, and eventually household tasks.
The company is taking a similar approach to data.
Sharpa collects training data from first-person human video, simulation, and real-world teleoperation. It says commercial settings and homes share some common objects, viewpoints, and manipulation patterns, potentially allowing data and learned skills to transfer between environments.
Continuous operation in commercial locations could also generate additional real-world data for further model development.
Whether those capabilities can transfer reliably beyond a controlled set of workflows remains a central question for general-purpose robotics. But for now, the DQ project gives Sharpa a commercial environment in which to test that premise.
The company sees food service as an initial market, followed by other service applications and, eventually, household use. More broadly, it expects hardware, data, and AI models developed for one environment to become increasingly reusable across others.
This article was adapted based on a feature originally written by Stone Jin and published on IPO Zaozhidao. KrASIA is authorized to translate, adapt, and publish its contents.
Note: RMB figures are converted to USD at rates of RMB 6.74 = USD 1 based on estimates as of September 3, 2026, unless otherwise stated. USD conversions are presented for ease of reference and may not fully match prevailing exchange rates.