A Robot Hand Needed 13,000 Simulated Years to Learn the Rubik's Cube, and in 2026 a Five-Finger Hand Still Screws In a Light Bulb Only 36 Percent of the Time

OpenAI’s robot hand solved the hardest Rubik’s Cube scramble in 2 of 10 attempts with its best setup, after the equivalent of roughly 13,000 years of practice in simulation. The paper reporting it went onto arXiv on 16 October 2019. Since then labs have taught robot hands with human demonstrations, touch sensors and cheaper hardware as well, and the newest measured results for a five-finger hand, from Google DeepMind on 30 July 2026, run from 32 to 92 percent depending on the task.

Every figure below is the one the named lab reported in its own paper or announcement. Most of them come from different robots, different objects and different definitions of success, so they are not a league table.

What OpenAI’s cube hand actually did

The robot was a Shadow Dexterous E Series Hand, a five-fingered humanoid hand, with three RGB cameras watching the cube. OpenAI trained the control policy and the vision system entirely in simulation and then moved them onto the real hand. The technique it credits is automatic domain randomization: the simulator keeps generating harder and harder variations of the world, so the policy cannot rely on any one version of it.

The paper puts the cumulative training experience for the cube policy at roughly 13 thousand years, running on 64 NVIDIA V100 graphics chips and 920 worker machines of 32 processor cores each.

The real-robot results sit in a table of ten trials per policy. The best policy, with the cube’s face angles supplied by a Giiker sensor, solved a scramble needing 15 face rotations in 60 percent of trials and the worst-case scramble of 26 rotations in 20 percent. Using vision alone to read the face angles, the same policy managed 20 percent and 0 percent. A trial ended if the cube was dropped or if the hand had not finished within 1,600 timesteps, about 128 seconds. The paper reports that the policy usually recovered when it turned the wrong face or the cube slipped, by regrasping.

Learning from a person’s hands

Simulation is not the only way to teach a hand. A person can show the robot the task, and the robot copies the movement. In April 2023 Tony Zhao, Vikash Kumar, Sergey Levine and Chelsea Finn published ALOHA, a low-cost bimanual (two-armed) system driven by a custom teleoperation interface, meaning a person drives the robot remotely. Their algorithm, Action Chunking with Transformers, learned six real-world tasks, including opening a translucent condiment cup and slotting a battery, at 80 to 90 percent success with about 10 minutes of demonstrations. The project page says the hardware was built with a $20,000 budget.

Pooling many labs’ demonstrations came next. The Open X-Embodiment paper, posted in October 2023, combined data from 22 robots across 21 institutions, covering 527 skills. On emergent-skill tests its RT-2-X model scored 75.8 percent against 27.3 percent for RT-2. The authors say they did not test robots with very different sensing and actuation, and did not study generalisation to new robots.

Cui and colleagues wrote in October 2025 that manual teleoperation of a dexterous arm and hand overloads human operators. Their fix splits the job: a person guides the arm through virtual reality while a learned policy drives the fingers using touch and camera feedback. They report 90 percent success across diverse objects, unseen ones included. In June 2025 Elvis Nava, Robert Katzschmann and colleagues described mimic-one, a 16-degree-of-freedom hand trained from glove and VR demonstrations, with up to 93.3 percent success on out-of-distribution tests, meaning situations that differed from the training data.

Robots can also learn from their own mistakes. Physical Intelligence’s pi*0.6 paper, posted on 18 November 2025, describes a training method called RECAP that mixes demonstrations, the robot’s own attempts and corrections from a human teleoperator. The company reports that the resulting model folds laundry in real homes, assembles boxes and makes espresso drinks, and that on some of the hardest tasks the method more than doubles throughput and roughly halves the failure rate. Laundry is also the chore tuput’s earlier piece on Moravec’s paradox used as its headline example.

Giving the fingertips a sense of touch

The F-TAC Hand, from a team including Zihang Zhao, Kaspar Althoefer and Song-Chun Zhu, carries tactile sensing at 0.1 millimetre spatial resolution across 70 percent of its surface. The preprint appeared on 19 December 2024 and Queen Mary University of London announced the published version, in Nature Machine Intelligence, on 9 June 2025. Across 600 real-world trials, the authors report, the tactile hand beat non-tactile alternatives on complex manipulation tasks (p below 0.0001).

Meta’s AI research group released the Digit 360 fingertip on 31 October 2024. Meta says it has over 8 million taxels (the touch-sensing pixels), picks up forces as small as 1 millinewton, and is paired with Sparsh, a touch model pretrained on over 460,000 tactile images. GelSight was to manufacture Digit 360, and Wonik Robotics a tactile Allegro Hand, both with wide availability planned for 2025.

Touch does not have to be elaborate. A Science Robotics paper by Qi Ye and colleagues, reported on 17 February 2026, used the low-cost four-fingered LEAP Hand with a standard webcam and basic touch or no-touch sensors, after pretraining on human demonstrations that combined vision and touch. The hand completed its eight tasks at a 73 percent success rate overall, and 85 percent on the five practised in simulation. The three unseen tasks, which included sharpening a pencil and unfastening a screw, were completed “most of the time” in the Phys.org report, which gives no figure. The listed failures include small bottle caps the hand lacks the torque to turn and fingers jamming in narrow faucet-handle slots.

Cheaper hands, and cables instead of motors

Kenneth Shaw, Ananye Agarwal and Deepak Pathak wrote in September 2023 that learning methods for dexterous manipulation had mostly stayed in simulation for lack of suitable hardware. Their LEAP Hand costs about $2,000 and takes four hours to assemble from readily available parts. The abstract puts that at about one-eighth the price of its closest competitor, the Allegro Hand.

Some newer hands move the fingers with cables, called tendons. mimic-one is one. So does Ruka-v2, from a team that includes Lerrel Pinto and Irmak Guzey, posted on 27 March 2026. It adds a two-degree-of-freedom wrist and sideways finger movement to the original Ruka, which the authors say could be built for under $1,300. In user studies on teleoperated tasks, Ruka-v2 cut completion time by 51.3 percent and raised success by 21.2 percent against the original. It was teleoperated across 13 tasks, and learned 3 of them autonomously.

Simulation came back

Simulation did not lose its place. Toru Lin, Jitendra Malik, Yuke Zhu and colleagues reported in February 2025 a reinforcement-learning pipeline trained in simulation for a Fourier GR1 humanoid with two multi-fingered hands. They report about 90 percent success on seen objects and 60 to 80 percent on novel ones. By task, the best policy averaged 62.3 percent for grasp-and-reach, 80 percent for lifting a box and 52.5 percent for a two-handed handover, at ten trials per object. The authors suspect the low handover figure comes from using only simple randomization of the simulated physics, and say the capabilities are far from general-purpose, human-level manipulation.

A February 2026 paper, DexRepNet++, reports 87.9 percent grasp success in simulation for a policy trained on 40 objects and tested on more than 5,000 unseen ones. Its abstract describes a small gap between simulation and real hardware but gives no real-world success rate.

DeepMind’s five-finger numbers from July 2026

On 30 July 2026 Google DeepMind announced Gemini Robotics 2. The model controls the five-fingered, 22-degree-of-freedom SharpaWave hand on Apptronik’s Apollo 2 humanoid. DeepMind’s post gives these success rates for that hand: unscrewing a light bulb 92 percent, screwing one in 36 percent, tying a trash bag 44 percent, a dustpan task 32 percent and a ziplock bag 40 percent. On two-finger grippers on a Franka Duo platform it reports 74.2 percent for general pick and place, 78.9 percent for tool kitting and 89.6 percent for precise insertion.

DeepMind’s post says multi-finger dexterous manipulation remains challenging. It also says new two-armed robots can typically be adapted with fewer than 200 examples.

Tuput’s report on humanoid pilots in 2026 covers the factory and airport side of these machines.

What still fails

Comparison is the first problem. A May 2026 survey of dexterous-hand research by Weiguang Zhao and colleagues says studies use different assumptions about robots, sensors, tasks, training data and evaluation, which makes systematic comparison hard.

The second is the gap between demos and unseen objects. The 2019 cube hand solved the worst-case scramble in 2 of 10 trials, Lin and colleagues’ novel-object figures trail their seen-object ones, and four of DeepMind’s five reported five-finger results are below 50 percent, all from DeepMind’s own tests. The Science Robotics team lists torque, jamming and slipping as open mechanical problems.

The third is access and speed. According to The Robot Report’s 2 August 2026 coverage, DeepMind limits the Gemini Robotics 2 model and its on-device version to early-access partners, and says robots have more to advance in movement speed.