Reduce one degree of freedom, Sharpa secures the right to define dexterous manipulation.

Sharpa has actively removed one degree of freedom on its new generation of dexterous hands.

On September 28, at the 2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) held in Pittsburgh, USA, Sharpa launched its new-generation dexterous hand W02. Compared with the previous generation product, it features a more compact structure and lighter weight, with tactile coverage expanded from the fingertips to the fingers and palm, while the number of active degrees of freedom is reduced from 22 to 21.

The degree of freedom determines how many types of movements the manipulator can control independently, and it is also one of the most intuitive parameters to understand the capabilities of a dexterous hand. For a company that once integrated 22 active degrees of freedom into a human-hand-sized structure, this trade-off is highly meaningful.

Sharpa new-generation dexterous hand W02

Launched alongside the W02 are the general-purpose humanoid robot D01 and the exoskeleton data glove AE01. The former provides full-machine motion and whole-body tactile perception, while the latter is used for teleoperation and data acquisition. Combined with the previously launched operation model CraftNet, Sharpa attempts to consider the design of the hand within the complete operation process covering perception, control and skill development.

By removing one degree of freedom, what exactly is lost, and what can be gained? This needs to be judged in the context of the tasks performed by the robot.

01. From 22 to 21, Sharpa rewrites the evaluation criteria for dexterous hands again

Degree of freedom is one of the most frequently compared indicators for dexterous hands. More independently controlled motion dimensions provide the manipulator with more posture and movement combinations, but also increase the complexity of structure and control. Whether they can be fully utilized depends on the requirements of tasks for the hand.

In May 2025, Sharpa launched its first-generation dexterous hand, integrating 22 active degrees of freedom, visio-tactile perception and human-hand configuration. This product, originally named Sharpa Wave and now called W01, has later been adopted by top universities around the world and NVIDIA's robot reference solutions. Being adopted by external R&D teams also makes Sharpa's product choices begin to influence how more groups implement dexterous manipulation.

When it comes to the second generation, the company began to re-examine the contribution of one of the joints.

The degree of freedom removed from W02 is the CMC (carpometacarpal joint) degree of freedom at the root of the little finger. Sharpa's preliminary application research found that this degree of freedom is used less frequently in common grasping and in-hand manipulation scenarios, which is mainly manifested as the little finger rotating along its own axis, and has limited contribution to most practical tasks. Eliminating it can simplify the structure of the little finger root and reduce potential failure points.

This adjustment first considers the balance between stability and the design of the entire hand. Whether adding one more motion dimension is worth the corresponding structure, weight and control costs needs to be judged in actual usage scenarios. Sharpa stated that the adjustment of W02 also incorporates feedback from exchanges with multiple customers.

Sharpa new-generation dexterous hand W02

"Less frequent use" is still limited to a specific task scope.

Low-frequency movements in common grasping may be useful in some special manipulation scenarios. To judge the cost of removing it, it is necessary to compare the completion status, failure causes and continuous operation performance of the two generations of products in the same task. Reducing a potential failure point is a consideration in structural design; how much the stability of the entire hand is improved still requires test results to verify.

Along with the reduction of degrees of freedom, changes in size and perception capabilities have also taken place. According to the company, the overall size of W02 is about 30% smaller than that of W01, and it is lighter in weight, which helps reduce the load on the forearm. The fingertips are equipped with high-resolution visio-tactile sensors, while the palm and other areas are covered by electronic skin.

Comparison between Sharpa dexterous hand W01 and W02

The significance of size is easier to see in the manipulation space. Door handles, tableware, tools, and a large number of work processes are all built around the size of human hands. An overly large mechanical structure, even with more motion capabilities, may be difficult to adapt to these environments.

The expansion of tactile coverage corresponds to another problem: the contact position between the object and the hand will change with the movement. When holding a soft object, the pressure needs to be adjusted according to deformation; when using a tool, it is necessary to take into account the finger position, contact state and movement trajectory. Where the contact occurs and whether the object slides will affect the next movement. The newly added perceptual information can only improve manipulation if it participates in the control in time.

With the change from 22 to 21 degrees of freedom, Sharpa puts more emphasis on stability, human-hand scale and contact perception while retaining high-degree-of-freedom motion capabilities. The complexity of the mechanical structure is not equivalent to the complexity of tasks that can be handled. Whether this trade-off is reasonable depends on whether the dexterous hand, after cooperating with the wrist and arm, can stably complete continuous tasks in constantly changing contact scenarios.

This also means that the competition for dexterous hands may be shifting from "parameter maximization" to "task contribution maximization". Degree of freedom, size, weight and sensing capabilities are no longer independent indicators. What is more important is whether they can be truly transformed into task capabilities under limited structure and control costs.

For the embodied intelligence industry, compared with "who builds more complex systems", "who can prove that every part of the complexity is useful" is the current vane.

02. After contact comes the tough part of dexterous manipulation

Previously, Sharpa's robots have been deployed in a DQ store in Shanghai, using the store's original equipment, tools and operation processes to participate in ice cream production. This type of task requires the robot to adapt to existing working procedures, and also puts the hand's motion capabilities, size and perception to the same test.

Taking stirring as an example, the robot needs to hold the paper cup while fixing the metal ring on the cup mouth with its thumb and index finger. The paper cup will deform, and stirring will change the force. If the grip is too loose, it is easy to lose stability, and if it is too tight, the cup may be crushed.

The difficulty here is to keep the fingers, wrist and arm cooperating continuously as the contact changes. The fingers maintain the position of the cup and the metal ring, the wrist adjusts the posture, the arm drives the cup to move, and the force feedback participates in controlling the grip strength.

Sharpa robot making ice cream in a DQ store

Sharpa's CraftNet processes motion planning and contact control in a layered manner.

System 1 is responsible for planning the overall movement and guiding the manipulator to approach the object; System 0 continuously fine-tunes the position of the hand and fingers based on tactile feedback at a frequency of about 100Hz. The execution state is then fed back to the upper layer to help the system correct the motion planning. On the robot, tactile feedback thus participates in grip control, enabling the robot to adjust its movements as the force changes.

This division of labor corresponds to different processing rhythms. The overall movement needs to advance according to the working procedure, but changes in contact may occur before the movement is completed. Layered control allows local adjustments to be performed more frequently.

Even if perception and control are connected, the robot may still make mistakes in the process of contact changes. The WM-CraftNet research from the Sharpa team, which was accepted by CoRL 2026, integrates wrist depth, tactile sensation, proprioception and motion history into a continuously updated state, trying to deal with real sensor noise as well as changes in object position and contact relationship.

In the W01 real-machine demonstration, the corner block can rotate continuously for more than one minute. In some disturbance tests, the system can also adjust the grip to restore rotation.

The research team also made public failure cases, for example, the duck-shaped object tipped over after completing five rotations. Continuously judging the contact state and restoring the grip in time after losing stability is still a difficult problem for the control system.

03. Defining the upper limit of dexterous manipulation depends on the entire system

This new product line continues the consideration of manipulation space and contact changes.

D01 adopts a spherical wrist joint close to the scale of the human body, which is convenient for adjusting the posture in a limited space; the electronic skin covering the arms, chest and other areas collects tactile information at a frequency of 100Hz. According to the company's disclosure, the maximum end effector speed of D01 exceeds 10.5m/s, and the repeated positioning accuracy is 0.2mm.

The size of the wrist is related to whether the hand can approach the object at an appropriate angle, while the body tactile sense extends the perception range beyond the fingertips. When the manipulated object touches the palm or the arm is subjected to external force, the control program also needs to judge how to respond to avoid local adjustments interrupting the ongoing task.

The way of applying force during contact also affects how the robot acquires manipulation skills. AE01 uses 22 encoders to capture the operator's natural joint movements and map the movements to the robot; each fingertip supports 256 levels of dynamic tactile feedback, which transmits the grip strength and contact state back to the human hand. Operators can adjust their movements according to contact feedback while seeing the object move, providing the teleoperation process with contact information corresponding to the movements.

The bartending demonstration at the IROS site adopts the teleoperation method, where the operator wears AE01 to control D01 to complete ice adding, wine pouring, shaking and serving in sequence. Another demonstration presents the robot's response to contact on the arms and chest.

Paper cup deformation and object sliding will change the operating conditions of the robot. Adjusting hardware, perception, control and data acquisition around the same task gives the team the opportunity to judge which link the problem lies in: whether the movement range is insufficient, the contact is not perceived in time, or the control fails to make effective adjustments.

If the real difficulty of complex manipulation is how to continuously perceive, judge and correct errors after contact occurs, dexterous hands can hardly be regarded as an isolated component anymore. The focus of the industry has begun to shift to the operating system composed of hands, body, tactile sense, control and data. Whether these capabilities can form a closed loop in the same task determines how complex real-world problems the robot can finally handle.

Existing models and task experience provide R&D foundation for new products, and the adaptation effect on D01 and W02 still needs to be verified through corresponding task tests.

04. Being the starting point for top teams' innovation gives the confidence to define the industry

The technical influence of a robotics company is also reflected in what conditions it can provide for others' innovation.

The selection experience of Nanyang Technological University illustrates how researchers weigh the capabilities of manipulators. In an interview published on Sharpa's official website, Assistant Professor Ziwei Wang introduced that the low-degree-of-freedom hands previously used by the team were limited in position accuracy and wrist-finger cooperation during screwing and assembly; another type of hand was too large to enter the assembly space of the gearbox. The final choice of W01 takes into account the accuracy, built-in tactile sensing, as well as supporting model files, deployment tools and teleoperation support.

He specifically mentioned the lateral movement of the metacarpophalangeal joint (MCP): "The key is to see which joint it is and what movement it can bring." This is a different movement capability from the little finger root CMC that W02 removed. For researchers, whether each joint is necessary depends on the task at hand.

After selecting the hardware, the team also needs to make these capabilities participate in control. The DexTeleop-0 research of Nanyang Technological University and other institutions builds a platform with two UR7e robotic arms and two W01 hands, adding tactile correction to human teleoperation commands.

7 operators conducted 5 tests for each method and each task respectively. In the Tube Operation task involving dual-arm cooperation and liquid transfer, the completion rate of the third stage increased from 34.29% when the movements were directly mapped to 77.14%. The hardware was not replaced, and the way tactile information was used changed the task results.

In the DexTeleop-0 research of Nanyang Technological University and other institutions, the ends of two UR7e robotic arms are equipped with two Sharpa Wave dexterous hands, which screw components into the gearbox and assemble smooth spheres onto the tubes

The same hand can also support different algorithm routes. The CAIP research of the University of California, Berkeley, NVIDIA and other institutions uses human videos and robot data to improve visual models, and W01