Robot hands: what humanoid hands can really do, and where they still fall short

Humanoid robot hands today can grasp a wide range of everyday objects, sense light touch and hold modest weights, but they remain far less capable than human hands. Product hands usually trade fingers and sensors for durability and cost, while research hands with many more joints show what is possible in controlled lab conditions. For events, that gap decides which hand tasks are realistic.

Why hands are the hard part

Walking robots get most of the attention, but manipulation is where many practical tasks succeed or fail. A hand has to position many small joints, apply the right force, adapt to objects of unknown shape, weight and texture, and survive repeated knocks. All of that has to fit inside a package light enough to sit at the end of an arm without overloading the shoulder and elbow motors.

The result is that manufacturers make clear compromises. Looking at the published specifications of several hands side by side shows those compromises more clearly than any demo video.

Degrees of freedom: what the number means

A degree of freedom (DoF) is one independent way a joint can move, such as a finger bending at its base or a thumb rotating across the palm. More degrees of freedom mean more possible hand shapes, but also more motors, more wiring, more weight and more things that can break.

There is also a difference between joints and actuated degrees of freedom. Many hands link several joints mechanically so that one motor curls a whole finger. This is called underactuation: the finger bends naturally around an object, but it cannot move each knuckle independently.

The published figures show a wide spread:

• Unitree Dex3-1, a hand Unitree pairs with its G1 humanoid: Unitree lists seven degrees of freedom across three fingers, with three in the thumb and two each in the index and middle fingers, a weight of 710 grams and a maximum payload of 500 grams when grasping a 5 cm round object.
• Inspire Robots RH56DFX, a five-finger hand: Inspire lists six degrees of freedom driving 12 joints, a weight of 540 grams, and fingertip forces of 15 N at the thumb and 10 N at each of the four fingers.
• Shadow Dexterous Hand, a long-established research hand: Shadow Robot lists 20 actuated degrees of freedom plus four underactuated movements, 24 joints in total, driven by tendons and 20 motors, with a weight of 4.3 kg.

The weight of the Shadow hand is the telling detail. Its dexterity comes at roughly six to eight times the mass of the compact product hands above, which is one reason research hands of this kind are usually mounted on fixed arms rather than walking robots.

Touch: the sense robots are still learning

People rely on touch to hold a paper cup without crushing it or to notice a glass slipping. Robots without touch sensing have to infer contact from motor current or from cameras, which is slower and less precise.

Tactile sensing in product hands has improved noticeably. Unitree lists 33 tactile sensors across the palm and fingertips of the Dex3-1, and its G1 page describes the tactile arrays as an optional fit. Inspire lists force sensors with a resolution of 0.5 N in the RH56DFX, which is useful for controlling grip strength but is far coarser than human touch.

When Figure AI announced its Figure 03 humanoid on 9 October 2025, it said it had developed its own fingertip tactile sensors able to detect forces as small as three grams, and that each hand includes a palm camera so the robot can still see an object when its head cameras are blocked. Figure also describes softer fingertips designed to increase contact area for more stable grasps. These are the company's own claims; independent measurements have not been published.

Research sensors go further. In October 2024 Meta's research division presented Digit 360, a finger-shaped tactile sensor that Meta says can detect forces down to 1 millinewton and also senses vibration and heat. Meta said GelSight would manufacture it and that Wonik Robotics would build a next-generation Allegro research hand integrating its tactile system. Sensors of this kind are aimed at researchers, not at event robots.

Software matters as much as hardware

A hand with many joints is only useful if software can coordinate them. For years, each new grasp needed careful engineering. Learning-based methods are changing that.

An early landmark was OpenAI's 2019 paper Solving Rubik's Cube with a Robot Hand, in which a control policy trained only in simulation manipulated a Rubik's Cube on a real humanoid robot hand. The authors credited a technique that steadily randomised the simulated environment so the policy would cope with the messier real world.

More recently, vision-language-action models aim to let one system handle many objects from plain instructions. In February 2025 Figure presented Helix, which it says controls the full upper body, including individual finger flexion and spreading, and lets its robots pick up thousands of household items they had not seen in training when asked in plain language. Figure's examples ranged from glassware to clothing. As with any company demonstration, the conditions were chosen by the company, and success rates across arbitrary real-world settings were not the focus of the announcement.

The gap between lab dexterity and product hands

Three gaps stand out when you compare research results with the hands on robots that can actually be hired.

1. Durability. Research hands are maintained by engineers between experiments. A hire robot may shake hundreds of hands, be bumped by visitors and be packed into a flight case at the end of the day. Simpler hands with fewer joints tolerate this better.
2. Strength and payload. The published payloads of compact product hands are measured in hundreds of grams per hand, not kilograms. Carrying a tray of full glasses or a heavy box is a different task from picking up a badge or a can.
3. Autonomy. Many impressive hand demonstrations are teleoperated, or run tasks that were trained in advance for specific objects. A robot on an event floor that is asked to handle an unfamiliar object without preparation is much less predictable.

There is also a cost trade-off. Each extra motor and sensor array adds price and maintenance, which is why many humanoids are sold with simple hands or no hands at all as standard, and with dexterous hands as an option. Unitree's G1 page, for example, lists the force-controlled three-finger hand as an option on the EDU version, while the standard G1 has no hand degrees of freedom.

What hands can realistically do at an event today

Based on the published specifications, the following hand tasks are within reach of many current humanoids when prepared in advance with the supplier:

• waving, pointing, gestures and choreographed hand movements;
• handshakes and high fives, within the force limits the supplier sets;
• picking up and handing over light, rigid, consistently shaped items such as a card, a small gift box or a can;
• holding a branded object or a sign for photos.

Tasks that remain risky without extensive testing include pouring drinks, handling fragile glassware, picking items from a mixed pile, and anything involving heavy objects or hot surfaces. Whether a hand interaction is autonomous or remote-controlled also changes what is reliable on the day, as explained in our guide on teleoperated and autonomous robots.

Physical contact with visitors also needs to be planned for. Grip forces of several newtons are small for a person, but pinch points in fingers and sudden arm movements are real hazards, so hand interactions belong in the risk assessment covered in our safety planning guide for public events.

What this means for event organisers

• Ask which hand is fitted. The same humanoid model can come with no hands, simple grippers or a dexterous hand. Check the exact configuration, for example on the Unitree G1 EDU page, and name it in your brief.
• Describe the object, not just the task. Size, weight, shape and fragility of anything the robot should hold matter more than the wording of the idea. Our guide on what to put in a request to a robot supplier lists the details suppliers need.
• Allow time for rehearsal. Handing over objects reliably usually means programming and testing with the actual items, so build it into the lead time discussed in custom programming and choreography.
• Treat demo videos with care. Company footage shows what a hand can do under chosen conditions. Ask the supplier whether a hand task has been performed at a live event before, and how it was operated.
• Keep a fallback. If a grasp fails, a gesture, a spoken line or a staff member handing over the item keeps the moment intact.

Frequently asked questions

How many degrees of freedom does a humanoid robot hand have?

It varies widely. Unitree lists seven degrees of freedom for its three-finger Dex3-1 hand and Inspire lists six for its five-finger RH56DFX, which uses linkages to move 12 joints. The Shadow Dexterous Hand, a research hand, has 20 actuated degrees of freedom across 24 joints. More degrees of freedom allow more hand shapes but add weight, cost and maintenance.

Can robot hands feel what they touch?

Some can, to a limited degree. Unitree lists 33 tactile sensors for its Dex3-1 hand, and Figure says the Figure 03 fingertips detect forces as small as three grams. Many hands instead estimate contact from motor force sensors, which is coarser. Research sensors such as Meta's Digit 360 go much further but are aimed at laboratories rather than event robots.

Can a hired humanoid robot hand out drinks or gifts at an event?

Handing over light, rigid and consistently shaped items such as cards, cans or small boxes is often feasible after rehearsal with the actual objects. Pouring drinks, carrying fragile glassware or heavy trays is much less reliable. Discuss the exact objects with the supplier, confirm which hand is fitted and plan a fallback if a grasp fails.

Why are lab robot hands more dexterous than commercial ones?

Research hands can use many motors, tendons and sensors because engineers maintain them between experiments and they often sit on fixed arms. The Shadow Dexterous Hand, for example, weighs 4.3 kg. Product hands for walking humanoids must be light, robust and affordable, so manufacturers use fewer actuated joints and simpler sensing.

How this article was made: written by the humanoid4rent editorial team with AI assistance; every factual statement was checked against the sources listed below on 2026-10-31. Spotted an error? Write to info@humanoid4rent.com.