Da Shen 1: China's Engineering Answer to the Four-Finger Dexterous Hand

“

Four-finger configuration convergence, task-driven selection, and hand-in variable configuration operation—In the global competition for embodied intelligent end-effectors, Chinese teams have already quietly submitted a pragmatic engineering answer.

Recently, Boston Dynamics (Boston Dynamics) published

specialized

New

A new four-finger dexterous hand developed by Atlas humanoid robot

，renewed the debate over the end-effector path. Compared to the common human-like five-finger scheme in the industry, the看点 of this hand is not just

The end effector should pursue complex human-like five fingers, or return to high rigidity

、high load

of extreme minimal configuration?

It is worth noting that when the spotlight focuses on Atlas's newcomers,

the global technology roadmap is actually quietly converging

——a coming

from

The four-finger dexterous hand of the Chinese team

，

has already provided its own engineering solution:

September 4, 2026

domestic humanoid robot end-effector startup Blackman Tech collaborated with Joyrobot

，jointly released

industrial-grade

variable

structure dexterous hand “Da Sheng 1” (SG100)

; while Boston Dynamics announced the release date of its four-finger hand as October 1, 2026. From the publicly released

schedule,

Da Sheng 1 comes first, with a 27-day difference

。

(see

Robotics Academy report

With Boston Dynamics official blog).

The order of release is just an objective fact, placing the two four-finger hands from across the ocean side by side for observation, more worth investigating is that, under the traction of industrial scenarios,

highly convergent engineering design

choices and

varied focus on breaking through

。

01.

convergent four-finger configuration: from the obsession of anthropomorphism to engineering choices oriented towards tasks

Figure 1 External comparison. Left figure

is the Da Sheng 1 product

。The figure on the right is

officially shot by Boston Dynamics

。The two figures have different viewpoints and display ratios, and are not used for comparing actual sizes.

In the field of dexterous hand research, "more fingers are better" has long been proven to be an engineering trap.

Traditional five-finger dexterous hands

often require stacking 15~20 active degrees of freedom and dense linear drive transmission systems, not only making the end weight and volume easily exceed limits, but also having high structural brittleness and maintenance costs in harsh working conditions, and bringing an explosive action space dimension to the upper-level reinforcement learning.

While

in industrial manufacturing sites, the core demands have always been highly focused

：facing

Multi-variety, variable-specification

workpieces, can the end-effector quickly form a stable force closure, provide sufficient gripping stiffness, and have the necessary pose adjustment tolerance during assembly or pick-and-place operations. Every additional finger comes with a sharp increase in marginal costs of motor drives, reducers, sensors, and control computing power.

The key to the problem has never been "how many actions can be performed", but rather

"which degrees of freedom are truly necessary, which mechanical redundancies can be discarded, and whether a more reasonable configuration design can be used to achieve the required

The limited freedom

degree is converted into higher task coverage capability.

Fig. 2 Boston Dynamics' official demonstration of Atlas tool operation画面。

Image source: Boston Dynamics。

From this perspective, it's not difficult to understand why the two end-effectors break away from the traditional five-finger

Abandon the redundant grip of the ring finger or little finger, relying on the opposing thumb and

Main finger

Build a stable three or four-point multi-face contact,

Coupling the determinism of mechanical hardware with the flexibility of upper-level algorithms.

Compared to merely pursuing high biomimicry, this configuration

emphasizes more

contact stability, structural stiffness, drive efficiency, and controllability oriented towards actual tasks

But similar appearances do not necessarily mean the same technical solutions.

In terms of technical implementation paths, the two companies have their own focuses: Boston Dynamics' new robot has

13 degrees of freedom

focusing on high-bandwidth joint direct drive, strong backdrivability, high-fidelity tactile perception at the fingertips, and Sim-to-Real closed-loop transfer; while BlackMamba Technology and Leju Robotics' DaSheng 1 uses

active

self

degrees of freedom

，其重点不只是“用更少的手指完成更多动作”，而是

围绕工业任务进一步引入主动可变构能力：

通过手指布局、对指关系与包络形态的切换，让同一只手在精细对捏、大行程夹持、环抱抓取和多指灵巧操作之间转换，并为后续掌心相机、触觉感知等模块接入预留进一步拓展空间。

这正是大圣 1 与普通四指灵巧手之间最值得关注的区别。四指是它的形态外貌，而“可变构”（Reconfigurable Kinematics）才是其工程哲学的底色。

02.

大圣 1 的解题路径：以“主动可变构”重构工业作业边界

传统机器人作业面临的最大瓶颈在于：工件千差万别，固定构型的末端往往只能在“灵巧但无力”或“有力但不灵”之间二选一。黑漫科技选择的破局点，不是让工件去适应手，而是让末端根据任务动态重构几何形态与接触拓扑——重新定义末端执行器与任务之间的关系：

不是为每一种工件准备一只不同的手，而是让同一只手主动适应不同的工件。

图3 大圣1多种异构手势及变换能力

针对产线复杂的作业包线，大圣 1 实现了四种典型构型的主动切换：

灵巧操作模式：

手指展开形成灵巧操作构型，满足多指协作与细小零件的精细触取；

对指平板模式：

主指与拇指形成高刚度对捏，专为精密插拔与高重复定位精度任务设计；

工业大行程模式：

扩张两组手指的有效跨度，胜任箱体、大直径圆柱等大尺寸物料的平稳转运；

环抱抓取模式

：

四指形成大包络抱紧，提供充裕的负载抓取力与抗冲击稳定性。

图4 大圣 1 产品资料展示的四种构型。不同布局对应不同的接触和夹持需求，图中为静态构型展示。

来源：黑漫科技产品资料。

对于需要面对多品种、小批量和柔性制造需求的工业场景而言，这种能力的价值并不只是增加一个抓取动作，而是

有机会延展单台机器人的工序覆盖度

，避免传统方案中频繁更换末端快换夹具带来的周期损耗与系统复杂性。

图5 乐聚机器人抓取实拍。图片展示了大圣 1 与工件的接触抓持场景。

来源：黑漫科技产品资料。

这种整机协同，正是大圣 1 与乐聚机器人深度合作的意义所在。作为机器人末端核心部件，大圣 1 并非脱离整机平台独立设计——黑漫科技与乐聚机器人的协同

打通了“末端核心零部件”与“人形机器人整机平台”的集成闭环

，手指的运动及动力学分布、通信响应延迟、动态延迟与结构强度等，得以直接置身于整机动态作业与产线工序中进行真实工况锤炼。

03.

跃迁手内操作：从固定抓持动作到整手动态重定向

在具身智能的操作序列中，“抓得起”只是第一

步，

“在手内调得动”

才是迈向高级灵巧作业的分水岭。

传统夹爪完成抓取后，物体位姿即被固化；但在

Precise assembly, work

Operational

in tasks such as object reorientation, etc.

，机器人必须具备在不松开物体的情况下，在掌内连续调整其姿态的能力——这就是In-Hand Manipulation，

In-Hand Manipulation,

“Hand-in-hand operation”

。

It requires the robot not only to stably grasp an object, but also to continuously adjust the contact relationship between the fingers and the object, enabling the object to rotate, reposition, and undergo continuous posture changes within the palm. This means the boundary of the end-effector's capabilities begins to change:

From 'picking up an object', moving towards 'operating an object within the hand'.

Around this direction, Da Shen 1 has built a set of scenarios for typical hand-in-hand operations

Reinforcement Learning Policy Library

, and completed multi-dimensional verification in high-fidelity physics simulation.

Figure 6 Da Shen 1 In-Hand Operation Reinforcement Learning Policy Simulation

(Top-left: Block Target Pose Redirect; Top-right: Single Ball Multi-Axis Continuous Rotation; Bottom-left: Dual Ball In-Hand Alternating Rotation; Bottom-right: Handheld Industrial Drill Stable Trigger Control)

Based on the inclusion of block redirect, single ball multi-axis rotation, and classic dual ball rotation tasks, Da Shen 1 focuses on verifying

continuous operation of handheld industrial drill trigger (In-Hand Drill Triggering)

— the end-effector not only needs to form a stable static support envelope using three fingers and the palm, but also requires an independent finger to apply high-dynamic, precise stroke triggering. This verification not only directly addresses the core pain points of human-robot collaborative tool use in industrial manufacturing, but also directly corresponds to the tool operation scenarios demonstrated by Boston Dynamics:

the four-finger configuration, after discarding one redundant finger, not only completely retains the ability to use complex tools, but also enhances contact stiffness and action determinacy through a more compact kinematic chain.

They demonstrate another layer of capability: when mechanical structures can provide sufficiently rich contact relationships, can robots further utilize these degrees of freedom to perform continuous operations on objects within the hand. This is particularly important for embodied intelligence—

large models and reinforcement learning can endow robots with stronger task planning and motion control capabilities, but ultimately, they still need to interact with the real world through the end-effector.

04.

From Simulation to Production Line: High-Fidelity Teleoperation and Data Closed-Loop in Real-Scenarios

The generalization of intra-hand operation strategies depends on the underlying hardware's precise mapping of real physical contact. Large models and embodied algorithms can plan action paths, but the data quality and success or failure of implementation depend on contact feedback and execution accuracy in real-world conditions.

To establish an end-to-end data collection and full-system verification closed-loop, DaSheng 1 built

low latency

delay,

high-fidelity teleoperation adaptation system

. In industrial sorting and loading/unloading production line scenarios, the Leju Kuafu full system platform equipped with DaSheng 1 conducted continuous high-dynamic material handling and sorting tests.

Figure 7 Real-time Dual-Arm Dynamic Production Line Material Grasping Test Based on Teleoperation

(covering multi-shape, variable-stiffness materials such as snack packaging bags, paper boxes, plastic cups, raw eggs, and medicine bottles)

In the dense production line rhythm, the operator drove DaSheng 1

to continuously handle targets with significantly different shapes and materials

: from easily slipping plastic cups and rigid paper boxes, to soft food packaging that is prone to deformation, and even highly fragile raw eggs.

This marks that DaSheng 1 has

crossed the stage of 'single-point hardware prototype'

By establishing a high-quality teaching data channel with low latency teleoperation, not only does it provide dense contact dynamics data for offline training of end-to-end policy models, but it also verifies the engineering reliability of directly embedding it into stringent working conditions such as industrial logistics and assembly lines.

If the model determines what the robot 'wants to do', then the end-effector determines whether the robot 'can actually do it'. From this perspective, the competition among dexterous hands will not only stay on the number of degrees of freedom, single-finger strength, or the degree of biomimicry, but will gradually move into a more practical dimension:

How many real-world tasks can a hand actually complete under limited mechanical complexity.

And the exploration of Da Sheng 1 from 'active variable structure' to 'in-hand operation' is exactly a technical path that follows the same logic progressively—first making the hand capable of actively adapting to different tasks, and then letting it continue to perform complex operations within the hand.

05.

27 days earlier, the engineering answer of the four-finger hand in China

The release 27 days earlier does not necessarily mean that the technological competition has already determined a winner, but it can clearly demonstrate:

In the current context of embodied intelligent hardware moving into deep water areas of scenarios, the Chinese robot industry chain not only possesses sharp industry sensitivity, but also has

engineering practice capabilities to resonate with global top competitors at key technology intersections, and even to lead in implementation.

When Atlas sparked a technological storm with the four-finger dexterous hand, looking back at this Chinese 'Da Sheng 1' that arrived 27 days earlier, we see not only a similar four-finger outline, but also an independent problem-solving and pragmatic exploration regarding industrial real-world pain points.

END

Industrial Robot Companies

Eston Automation

Efte Robot

Fao Robot

Yuejiang Robot

Jieka Robot

Songling Robot

Luoshi Robot

Atongmu Robot

Jizhi Jia

Hikvision Robot

Yifei Technology

Ailite Robot

Service and Specialized Robot Companies

Yijiahe

Jingpin Tieshang

Qiteng Robot

Shihe Robot

Pudu Robot

Shiloude Robot

Kuma Technology MAMMOTION

Humanoid Robot Companies

Ubot Technology

Yushu

Yundeshen

Xingdong Jiyuan

Weijing Robot

Zhujidongli

Leju Robot

Daxiang Robot

Magic Atom

CrowdBot Robotics

PASINI Perception

Cyborg Robotics

Digital China

Fugui Intelligent

Tianlian Robotics

Kepler Humanoid Robotics

Lingbao CASBOT

Qingbao Robotics

Zhejiang Humanoid Robotics Innovation Center

Dongyi Technology

Zhishen Technology

PNDbotics

Zhuoyide Robotics

Qinglang Intelligent

Galileo GALILEO

Songyan Power

Tianji Intelligent

Woaan Robotics

LiGong Huawei

Jiace Evolution

Embodied Intelligence Enterprises

Cross-Dimensional Intelligent

Galaxy Universal

Qianxun Intelligent

Lingxin Qiaoshou

Ruierman Intelligent

Weiyi Intelligent Manufacturing

Tuisheng Technology

Zhongke Silicon Chronicle

Shu途 Technology

Lingqiao Intelligent

Xingchen Intelligent

Qiongche Intelligent

Fangzhou Infinite

| iFlytek |

Beijing Humanoid Robotics Innovation Center

National and Local Co-construction Humanoid Robotics Innovation Center

Daimeng Robotics

VisiBots Robotics

Star Sea Chart

Moon Spring Bionic

Zeroth Square Robot

Zhongke ShenGu

Smart Square

Big Name Robot

Hao Cun Technology

Ju Shi Intelligence

Xynova Xi Nuo Future

Fei Xi Technology

Future Power

Boden Intelligence

Qian Jue Technology

Ling Sheng Technology

Ji Cui Intelligent Manufacturing

Xin Bait Technology

Chen Hun Line Technology

Dexmal Original Power

You Li Qi

Self Variable

Ruiyan Intelligent Control Dexterous Hand

Qi Wu Technology

RoboScience Robot Science

Zhongke Fifth Dynasty

Critical Point

Dang Hong Technology

Qiao Jie Shu Wu

Vbot Vita Power

He Shan Technology

Ju Brain磐 Stone

You Ai Zhi He Robot

Zhi Xing Jian

A Miao Robot

Lu Ming Robot

Quantum Intelligence

Ru Yu Technology

Tie Hu Robot

Zhi Wu Ji

Medical Robot Enterprise

Yuan Hua Intelligence

Tian Zhi Hang

Si Zhe Rui Intelligent Medical Care

Jing Feng Medical

Tuo Dao Medical

Zhen Yi Da

Shu Rui® Robot

Luosen Bot

Shui Mu Dong Fang

｜

Kang Nuo Siten

Di Shi Medical

Upstream Industry Chain Enterprise

Lv De He Bo

Yin Shi Robot

Kun Wei Technology

Mai Ta Intelligent

Qing Tong Vision

Ben Mo Technology

Lan Dian Touch Control

| Xin Jing Cheng Sensor

BrainCo Qiang Brain Technology

Yu Li Instrument

Ji Ya Jing Ji

Si Lan Technology

Shen Yuan Sheng

Fei Pu Navigation Technology

Inks

Juxia Intelligent Drive

Ling Yun Guang Yuan Ke Vision

Xuan Ji Power

Yi You Technology

| Rui Yuan Precision |

Ling Zu Era

HIT Hua Wei Ke

Xing Hui Sensor

Ling Di Technology

Quan Zhi Bo

CubeMars Robot Power

Wang Long Robot Elevator

Hang Kai Micro Electronics

Wu Tong Sensing Control