DexBench: Robot Hand Skills Home Humanoids Need

That is why DexBench is interesting for home-robot buyers. Built around industrial dexterity tasks, it gives us a better vocabulary for asking what a robot hand can actually do: grasp different shapes, place contacts precisely, time motion, regulate force, and recover when the task changes halfway through.

The short version: DexBench is not a shopping scorecard yet. It is a benchmark language. But if you are comparing home humanoids, assistive robots, or mobile manipulators, it points to the questions that matter more than "how many fingers does it have?"

What is DexBench?

DexBench describes itself as an industry-grade benchmark for robotic dexterity. Its own framing is blunt: manipulation research has focused heavily on grasping, but industry needs much more than grasping. The site currently lists 18 benchmark tasks and 56 evaluation cases, organized around object-state complexity and five dexterity regimes.

The companion RLDX-1 technical report from RLWRLD matters because it explains why the benchmark exists. RLWRLD says frontier vision-language-action models often fail on tasks such as pouring, in-hand rotation, and reactive grasping because the missing information is not just visual. The model needs high-DoF hand-object geometry, temporal dynamics, contact forces, torque, tactile feedback, and memory.

That maps closely to the home. A useful home robot may need to pull a towel from a pile, rotate a bottle to read the label, hand over a plate without squeezing too hard, pick a dropped charger cable, or stop pouring when the mug is full. Those are not one generic "pick up object" skill.

DexBench's useful buyer insight is that dexterity is not a property of the hand alone. A two-finger gripper with the right sensing and control can be more competent for a narrow task than a five-finger hand driven by a weak policy. A humanoid hand should earn trust by passing hard task regimes, not by looking human in a launch video.

The five skills buyers should ask vendors to prove

DexBench organizes dexterity into five regimes. Translated into home-robot language, they become a practical checklist.

This framing is better than a spec-sheet race. A robot with a 20 kg payload can still be bad at threading a small cap. A robot with elegant fingers can still fail if it has no tactile feedback. A mobile base can carry the arm to the counter, but the hand still needs to solve contact.

The RLDX-1 project page makes the same point from the model side. It describes separate components for motion awareness, physical sensing, and memory, rather than assuming that a bigger VLA will magically infer contact forces from video. Its public GitHub page says the open checkpoints include pre-trained and mid-trained models in the 7B to 8B parameter range, and the arXiv abstract reports RLDX-1 outperforming recent VLA baselines on simulation and real-world dexterous-manipulation tasks.

For buyers, the exact benchmark leaderboard is less important than the principle: the vendor should explain which failure modes the robot can handle. If the answer is only "AI," treat that as a warning sign.

How current home-robot hardware maps to DexBench

ui44's database already shows why the benchmark is useful. The most credible home-adjacent robots have very different manipulation stacks.

Reachy 2 is not a consumer home robot, but it is a useful reference point because it is built for research manipulation. ui44 tracks it at $70,000, with 3 kg per arm, ROS 2, a Python SDK, and compatibility with Hugging Face LeRobot. That makes it relevant to developers who want to train and test manipulation policies, not to ordinary buyers looking for a finished chore robot.

Figure 03 is the opposite kind of signal: a full humanoid with no public price yet, about 5 hours of battery life, a 20 kg listed payload, Helix VLA, force sensors, and tactile arrays in the ui44 record. Those sensors are exactly the kind of hardware contact-precision tasks need, but the buyer question is proof: which household tasks have been repeated, measured, and published beyond short videos?

1X NEO is tracked as a $20,000 pre-order humanoid with a 30 kg body, around 4 hours of battery life, depth sensors, tactile skin, and 1X embodied intelligence. NEO is one of the more directly home-positioned humanoids in the database, so DexBench-style questions are especially relevant: can it recover from a failed grasp, or does a remote operator take over when the object moves?

Hello Robot Stretch 4 is a reminder that home manipulation does not require a biped. It costs $29,950, uses a mobile base, has a wrist-mounted depth camera, and ui44 tracks its arm payload at 2.5 kg extended or 4 kg retracted. Its gripper is not anthropomorphic, but for assistive pick-and-place, clear sensing and a bounded workspace can be more important than looking human.

Unitree G1 starts at $13,500 and brings humanoid walking into a lower price band, but ui44 tracks arm payload around 2 kg on the standard model and about 3 kg on G1 EDU. NEURA 4NE-1 Mini starts at €19,999 before taxes/shipping, with a €29,999 Pro tier that adds 12-DOF dexterous hands, SDK, digital twin, and teleoperation. LimX Oli lists a 3 kg single-arm load and contact-sales pricing. These are very different buyer propositions, but DexBench asks the same thing of all of them: which dexterity regimes are proven, and under what conditions?

What DexBench does not prove for a home buyer

DexBench is valuable, but it is not a consumer certification label. Most of its current examples are industrial: insertion, fastening, bimanual regrasping, cutting, cable handling, and other tasks that stress manipulation. Those tasks are relevant because homes also contain deformable, slippery, occluded, and awkward objects. But a benchmark pass in a lab does not prove that a robot can work safely in your kitchen.

Three gaps matter.

First, homes are messier than benchmark scenes. The same spoon may be in a sink, under a napkin, next to a glass, or partly wet. A benchmark can isolate a skill, but a home robot has to decide when not to attempt it.

Second, safety is broader than manipulation success. A hand that can grip firmly also needs force limits around pets, children, glassware, skin, and electronics. Contact precision should include the ability to stop, release, and ask for help.

Third, autonomy and teleoperation often blur together. A robot may have excellent hardware, but if the impressive household demo depends on a remote human supervisor, buyers need to know that. Teleoperation can be a sensible transition strategy, especially for assistive care, but it is not the same as a fully autonomous chore robot.

That is why the right question is not "did the robot pass DexBench?" The better question is: which benchmark-like claims are backed by repeated trials, clear success criteria, and disclosed limits?

Questions to ask before believing a robot-hand demo

When a vendor shows a home humanoid manipulating objects, ask for specifics. Good answers do not have to be academic papers, but they should be concrete.

1. What objects were tested? Thin, soft, reflective, transparent, wet, and deformable objects are much harder than large rigid props.
2. What changed between trials? A true grasp-diversity claim should vary size, orientation, lighting, clutter, and starting pose.
3. Was contact sensed or guessed? Force/torque and tactile signals matter for pouring, tightening, handovers, and fragile objects.
4. How often did it fail? A single take is not a success rate. Ask for the denominator.
5. What happens after failure? Context awareness means the robot can retry, regrasp, ask for help, or abandon safely.
6. Was a person controlling it? Teleoperation, autonomy, and supervised autonomy should be labeled separately.

Those questions also help distinguish robot categories. A mobile manipulator like Stretch can be honest and useful with a narrow task envelope. A consumer humanoid needs a broader claim because it is asking for more trust, more money, and more space in the home.

Should home-robot buyers care about DexBench now?

Yes, but as a reality filter, not as a buying requirement.

DexBench gives buyers and developers a vocabulary for the next phase of home robots. The first phase was mobility: can the robot navigate without crashing? The current phase is perception and language: can it understand a scene and a spoken instruction? The harder phase is contact: can it change the world safely with its hands?

That is where the home-robot market will separate. Social robots and camera robots can be useful without dexterous manipulation. But any robot claiming to clean up, cook, fold laundry, care for a person, or load a dishwasher needs to prove hand skills across grasp diversity, spatial precision, timing, force, and recovery.

For now, DexBench is best used as a checklist. When a company says its humanoid has human-like hands, ask which DexBench-style regimes it has actually demonstrated. When a company says its robot has a powerful AI brain, ask whether that brain sees motion, remembers task progress, and senses contact. And when a demo looks too smooth, ask what the robot did after the first failed grasp.

The robots that answer those questions clearly will be more credible than the ones with the most cinematic hands.