Robotics 5 November 2026 4 MIN READ
How a Robot Hand Grasps: Touch, Data and Simulation
How robot hands learn to grasp: demonstrations, simulation and touch sensing, with the numbers from the papers and what Boston Dynamics says about Atlas.
THE SHORT ANSWER
A robot hand needs demonstrations from people, accurate simulation and a sense of touch. Published results show each alone has limits: a 2023 system used simple grippers, and a 2019 Rubik's cube hand completed the full scramble in only two of ten trials at best.
On this page
Why a hand is harder than an arm
Boston Dynamics, which has recently redesigned the hand of its Atlas humanoid, puts the problem bluntly: making a useful hand is “a ruthless tradeoff” between dexterity, strength, ruggedness, cost, repairability and sensing. Every finger adds joints, motors and things that can break. The company’s October 2026 post says a multi-fingered hand is “like a whole mini-robot”, and its new one has four fingers and 13 degrees of freedom, up from 7 in the previous generation.
The same post explains a design choice that shows how far function beats imitation. The new hand has no little finger. The team’s chief technology officer asked them to tape their little and ring fingers together for a day and report what they could not do. The conclusion was that a fifth finger would add three actuators, and with them cost, volume and failure risk, for little gain. This is the company’s own account, not an independent evaluation.
Teaching a hand: three routes
Researchers use three broad ways to get a hand to grasp well, and each has a clear limit.
1. Learning from a human operator. In 2023 a Stanford, Meta and Berkeley team published ALOHA, a bimanual teleoperation system built from off-the-shelf arms and 3D-printed parts for under 20,000 dollars. A person moves a pair of “leader” arms and two “follower” arms mirror them. From 50 demonstrations per task, about 10 to 20 minutes of data, their learning method opened a translucent condiment cup and slotted a battery with 80 to 90% success, according to the authors.
The detail worth noticing is the hardware. The paper states that the team did not use dexterous hands “due to price and maintenance considerations” and built the system around parallel-jaw grippers. The authors also list what failed: unwrapping a candy worked 0 times out of 10 in a preliminary test, and buttoning a shirt was beyond the system.
2. Learning from simulation. In 2019 OpenAI trained a Shadow Dexterous Hand, a five-fingered robot hand, to manipulate a Rubik’s cube using only simulated experience. The training used an automatic way of making the simulation harder as the policy improved, and about 13,000 years of cumulative simulated experience. The hardware needed more than a hand: three cameras, a motion-capture system for the fingertips and, for the best results, a cube with built-in sensors.
The results are more modest than the headline. In ten trials with the sensor cube, the best policy completed at least half of a fair scramble six times and the full scramble twice. With vision alone it completed the full scramble in none of ten trials. A policy trained with the earlier, manual method never completed even half.
3. Learning from touch. A hand that cannot feel has to guess whether an object is slipping. The GelSight sensor, developed at MIT, puts a soft elastomer pad over a small camera. The pad deforms against an object and the camera records the shape, from which contact force and slip can be inferred. Its review paper reports a spatial resolution of 30 to 100 microns for compact versions meant for robot fingers, against 5 mm for a commercial pressure array it cites. Boston Dynamics says its new hand has dense pressure sensors on the fingertips and palm, but the post gives no figures for them.
Where the debate is
Boston Dynamics argues that gloves and handheld devices are a good way to collect human demonstrations, but capture mostly “proxy signals” and not the fast force control a dexterous hand needs. It concludes that reinforcement learning in simulation is “an essential component”, and that the hand was built to be simulated accurately. The OpenAI result supports half of that: simulation can transfer to a real hand, but only with heavy randomisation and external sensing. The ALOHA result supports the other half: demonstrations can teach useful skills from little data, but on simple grippers.
What we cannot tell yet
For the new Atlas hand there are no independent measurements in the sources we read, and the company’s post gives no success rates, tactile sensor specifications or failure data. The OpenAI and ALOHA papers date from 2019 and 2023, and neither used the Atlas hand. Their numbers show what each method achieved on its own hardware and tasks and cannot be compared with one another.
Our take: a robot hand is a bet about which signals matter. Human demonstrations, simulated physics and touch each supply one of them, and each of the studies here works with only one or two. The sensible reading of the 2026 Atlas hand is as a hardware choice that makes all three easier to combine, whose payoff is still to be shown in a factory.
Questions and answers
Why do humanoid robots have so few fingers?
Each finger adds joints and motors that cost money and can break. Boston Dynamics says its new Atlas hand has four fingers and 13 degrees of freedom and no little finger, after a team test of living without one for a day.
What is a tactile sensor on a robot hand?
A sensor that tells the robot where and how hard it touches an object. MIT's GelSight uses a soft pad and a camera to record the shape of the contact, with a resolution of 30 to 100 microns on compact versions.
Did a robot hand really solve a Rubik's cube?
In 2019 OpenAI reported it did, using a five-fingered hand trained in simulation. In its own ten trials with a sensor-equipped cube, the best policy completed the full scramble twice, and none of ten with vision alone.
Can robots learn dexterous skills from human demonstrations?
Yes, within limits. A 2023 system learned tasks such as slotting a battery with 80 to 90% success from about 10 minutes of demonstrations, but it used parallel-jaw grippers, not dexterous hands, and failed at tasks like unwrapping a candy.
Do you agree?
Take it to Chevlume: write your view, add your sources and let others answer. The closed alpha is by invitation.
Keep reading
From BigDog to Atlas: Hydraulic to Electric
From the 2008 BigDog paper to a production electric Atlas in 2026: what Boston Dynamics and its sources say, the specifications and what is still unknown.
ROBOTICS / EXPLAINERHow Humanoid Robots Actually Work
Motors, sensors, balance and hands: what a humanoid robot is made of, what it does reliably today and why spec sheets flatter the machine.
ROBOTICS / DISCOVERYIngenuity: The Mars Helicopter
Planned for five flights, it made 72. How a helicopter with a mobile-phone processor flew in air 1% as dense as Earth's, and what ended its mission.
Sources
- Robot Hands for Modern AI and Real Work (1 October 2026) · Boston Dynamics · Accessed
- Solving Rubik's Cube with a Robot Hand (OpenAI et al., 2019) · arXiv / OpenAI · Accessed
- GelSight: High-Resolution Robot Tactile Sensors for Estimating Geometry and Force (Yuan, Dong, Adelson), Sensors 17(12), 2017 · MDPI Sensors via PubMed Central · Accessed
- Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (Zhao, Kumar, Levine, Finn, 2023) · arXiv / Stanford University · Accessed
Published
Discuss on X