Sharpa unveils humanoid D01, dexterous hand W02, and exoskeleton glove AE01 at IROS

What does it take for a robot to move beyond stage demos and create real commercial value? At the ongoing IROS conference, embodied AI company Sharpa offered its answer, unveiling three products at once. The company, founded by the team behind lidar maker Hesai, has long been known for high-degree-of-freedom tactile dexterous hands — and this time it put the hand, the body, and the data entry point on the same stand.

The first product is D01, described by Sharpa as an all-in-one tactile-perception dexterous manipulation robot and its first fully self-developed general-purpose humanoid. Its headline numbers all revolve around manipulation: an arm payload-to-weight ratio close to 1:1, a maximum end-effector speed above 10.5 m/s, a 1000 Hz communication rate, 0.2 mm repeat positioning accuracy, and a spherical wrist designed at a 1:1 human scale. High speed lets it perform dynamic tasks like passing a baton or handling tools, while 0.2 mm repeat accuracy pushes toward fine manipulation.

The more unusual part is touch. D01's electronic skin covers the whole body, with tactile coverage across the upper body, a 100 Hz tactile sampling rate, a force sensing range of 0.1–20 N, and 0.2 N force resolution, so it can directly sense light contact, collisions, and external forces. In dexterous manipulation, vision can tell a robot that there is an object ahead, but once contact happens the robot also needs to know where it touched, how much force is involved, and whether the contact is changing — transient physical events that vision alone struggles to capture.

Sharpa's approach is to fold this body feedback into the robot's perception system. With full tactile coverage across D01's core interaction areas, the body shifts from a channel for anomaly detection or after-the-fact logging into part of action adjustment. Just as a human subconsciously grips harder when a finger senses slipping, the robot needs the same real-time feedback.

The second product is W02, a next-generation fully tactile, ultra-compact, lightweight dexterous hand. One surprising change: the previous W01 had 22 active degrees of freedom, while W02 drops to 21. The removed degree sits at the base of the pinky (the CMC joint); according to Sharpa's application research, it is rarely used in common grasping and in-hand manipulation and contributes little to tasks, so simplifying it also cuts mechanical complexity and potential failure points. The lesson is that a dexterous hand does not have to keep increasing its degrees of freedom without limit.

W02 packs its 21 active degrees of freedom into a tighter structure, reducing overall hand size by roughly 30% versus W01 and lowering weight further — cutting forearm load and making it easier to fit into workspaces designed around human hand dimensions. Smaller and lighter also unlocks tasks that were previously hard: W02 achieves a zero grasp radius to handle small-diameter objects like chopsticks and thin wires more stably, with a high-resolution visual-tactile sensor at the fingertips covering a 5 mN–30 N force range at 1 mm spatial resolution, while electronic skin covers the rest of the palm.

The third product is AE01, 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 motions and map them to the robot in real time. The key is feedback: when a person handles an object, the eyes see position and shape while the fingers feel force, contact, and slip — AE01 sends that information back to the operator so they can adjust. It fills in the final piece on the data side, capturing human motion trajectories, the robot's contact feedback, and the mapping between them in a single collection process.

Read together, the three products show Sharpa trying to turn dexterous manipulation from isolated hardware capabilities into a system built around contact sensing, closed-loop control, and data learning. Its CraftNet model is split into three layers: System 2 handles task understanding and long-horizon planning, System 1 handles pre-contact motion planning, and System 0 enters the contact phase to process tactile feedback and fine motions at higher frequency, feeding state back to System 1. What touch really changes here is the timescale of the control loop.

Pushing further, Sharpa's latest WM-Craftnet research trains a World Synesthesia Model with temporal memory. Inputs include wrist depth, touch, proprioception, and previously executed actions, compressed through a Dreamer-style recurrent state-space model into a time-varying latent state that serves as policy context for judging an object's current geometry, contact, and motion. The model learns not only what is being touched now but why the state became what it is. The question ahead is whether this full stack can move from the IROS stand into real production lines and deliver on the promise of commercial value.