Sharpa Unveils Humanoid Robot D01, Dexterous Hand W02 and Data Glove AE01 at IROS

Sharpa describes D01 as an integrated tactile-sensing dexterous manipulation robot. Its arm payload-to-self-weight ratio is close to 1:1, maximum end-effector speed exceeds 10.5 m/s, communication frequency reaches 1000 Hz, repeat positioning accuracy is 0.2 mm, and its spherical wrist is designed at 1:1 human scale. The robot's electronic skin covers its entire body, with tactile sensing across the upper body; tactile sampling is 100 Hz, force sensing range is 0.1-20 N, and force resolution is 0.2 N. Sharpa says these specifications are intended to let the robot handle dynamic tasks such as baton passing and tool use while retaining precision and force margin.

QbitAI reported that Sharpa places tactile sensing at the center because vision can indicate that an object is ahead but cannot fully describe what happens after contact. A robot needs to know where contact occurred, how much force is applied and whether contact is changing when catching a falling object or adjusting a trajectory after a person touches its arm. D01's skin coverage extends this kind of bodily feedback into the robot's control system, turning it into part of motion adjustment rather than only anomaly detection or after-the-fact logging.

W02 is described as a new-generation full-tactile, ultra-compact and lightweight dexterous hand. It has 21 active degrees of freedom, one fewer than W01's 22. The removed degree of freedom is at the base CMC joint of the little finger. Sharpa's application research found that this degree of freedom is used infrequently in common grasping and in-hand manipulation, so removing it reduces mechanical complexity and potential failure points. The whole hand is about 30% smaller than W01 and lighter. W02 achieves zero grasp radius, which Sharpa says helps it handle small-diameter objects such as chopsticks and thin wires. Its fingertips use high-resolution visuotactile sensors with a force range of 5 mN-30 N and 1 mm spatial resolution; electronic skin on the rest of the palm covers 0.1-20 N with 5 mm spatial resolution.

AE01 is a high-fidelity exoskeleton haptic data glove for precise teleoperation and first-person data collection. It uses 22 encoders to capture an operator's natural hand movements and map them to the robot in real time. The glove also feeds tactile information back to the operator, so actions can be adjusted according to force, contact and slip. QbitAI reported that this creates a two-way interface: human motion enters the robot, while the robot's contact state affects the operator's next move. The data collected in one operation therefore includes not only hand trajectories but contact timing, force changes, slip states and correction actions. Sharpa says precise control, synchronized tactile collection and calibration-free operation reduce barriers to human-robot mapping and multimodal data synchronization.

At the control and model level, Sharpa's CraftNet uses System 2 for task understanding and long-horizon planning, System 1 for pre-contact motion planning, and System 0 for the contact phase, where it processes tactile feedback and fine motion at higher frequency and sends state information back to System 1. The company's WM-Craftnet research trains a World Synesthesia Model, a world model with temporal memory. Inputs include wrist depth, touch, proprioception and previous actions; a Dreamer-style recurrent state-space model compresses them into a time-varying latent state used as policy context to judge an object's geometry, contact and motion state. QbitAI reported that this moves beyond simple visual-tactile fusion by having the robot form an internal representation of contact state.

Sharpa's manipulation stack has also entered a commercial setting. In late August, a Blizzard robot restaurant opened on Wujiang Road in Shanghai in partnership with DQ. The robot did not receive a dedicated workstation; it uses DQ's existing equipment, ingredients and preparation process. It completes 55 consecutive steps, from opening cabinet doors, placing cup rings, dispensing ice, adding toppings and mixing to pouring. QbitAI reported that the difficulty lies in completing a highly correlated sequence without modifying the environment, where errors from one step can carry into the next.

Sharpa co-founder Li Yifan told QbitAI that “purchased data cannot build a moat.” The company's product layout reflects that view, QbitAI reported: rather than relying on a single model or data-collection method, Sharpa aims to generate action-contact-feedback-correction experience in real tasks and feed it back into models. After scaling autonomous-driving production, the founding team continues to emphasize commercial generalization in embodied AI, including whether a system can keep working after changing scenes or product categories and whether efficiency gains can cover procurement, deployment, maintenance and data costs.

QbitAI reported that Sharpa's earlier W01 dexterous hand addressed whether a robot can hold and manipulate objects. With D01, W02, AE01 and the DQ deployment, the company is pursuing a broader question: how to distill general hardware and technical requirements for different tasks and provide a general manipulation capability that can take real jobs, continuously deliver, accumulate data and improve.