Robots that use AI to handle objects learn mainly from one of two kinds of experience: simulated practice or recorded human demonstrations. Three published projects show how far each goes. OpenAI trained a robot hand in simulation and solved a Rubik's Cube on real hardware, Google's RT-1 learned from about 130,000 demonstrations and reached 97 percent success on its trained instructions, and the Open X-Embodiment collaboration pooled data from 22 robot types at 21 institutions.
Key takeaways
- Simulation lets a robot practice far more than real hardware allows, but the gap between simulated and real physics must be closed with randomization.
- The 2019 Rubik's Cube hand solved a cube needing 15 face rotations 60 percent of the time and the hardest, 26 rotations, 20 percent of the time.
- RT-1 reported 97 percent success on seen instructions and 76 percent on unseen ones, using 13 robots over 17 months.
- Open X-Embodiment pooled 22 robot types and reported positive transfer between them.
- Published success rates describe specific tasks and test setups, so they do not rank robots or show general ability.
Why "best robot" is the wrong question
Lists that name the best AI robots of a decade mix products built for different jobs: a warehouse arm, a humanoid for demonstrations, and a research hand cannot be scored on one scale. A more useful question is how a robot acquires a skill, because that method decides what it can do and where it fails. The projects below were chosen because their papers state the method, the amount of data, and the results, which allows the numbers to be checked.
Learning in simulation: the Rubik's Cube hand
In 2019, OpenAI researchers reported that a five-fingered robot hand had solved a Rubik's Cube after training only in simulation. The key technique was automatic domain randomization. The simulator changes properties such as friction, object size, and mass, and widens the range as the policy improves, so the real world looks like one more variation. In the paper's best configuration, which used a cube with built-in sensors, the hand averaged 26.8 successes over 10 trials, which the authors translate to a 60 percent success rate for a cube needing 15 face rotations and 20 percent for 26, the worst case. A vision-only version succeeded 20 percent of the time on the shorter scramble and not at all on the longest.
Learning from demonstrations: RT-1
RT-1, the Robotics Transformer from Google researchers, took the opposite route. Instead of simulation, it used real data: more than 130,000 demonstrations collected over 17 months with 13 robots, covering over 700 language instructions. The robots were mobile manipulators with a seven-degree-of-freedom arm and a two-fingered gripper. RT-1 reported 97 percent success on seen instructions and 76 percent on unseen ones. The cost is the data itself: every new skill needs people to demonstrate it on hardware, which is slow and expensive.
Pooling data across machines: Open X-Embodiment
Most robot datasets belong to one lab and one machine. The 2023 Open X-Embodiment collaboration combined data from 22 robot types contributed by 21 institutions, covering 527 skills and 160,266 tasks. The authors trained a model called RT-X and reported positive transfer, meaning that experience from other robots improved the capability of several of the robots. This is early evidence that data from one machine can help another, not proof that a single model controls any robot.
Three methods side by side
| Project | Data source | Reported result | Main limit |
|---|---|---|---|
| Rubik's Cube hand (2019) | Simulation with domain randomization | 60 percent at 15 rotations, 20 percent at 26 | Best results used a sensor-equipped cube |
| RT-1 (2022) | About 130,000 real demonstrations | 97 percent seen, 76 percent unseen instructions | Demonstrations must be recorded on real robots |
| Open X-Embodiment (2023) | Data from 22 robot types, 21 institutions | Positive transfer across robots | Gains vary by robot and task |
Checking a robot demonstration claim
- What exactly was the task, and is it the same task shown in the video?
- How many trials were run, and were failures reported?
- Did the robot use special sensors or marked objects that a real home or workplace would not have?
- Were the test instructions or objects new, or did they appear in training?
- Was a person resetting the scene between attempts?
FAQ
Why train a robot in simulation at all?
Simulation allows many practice attempts without wear on hardware or risk to people. Its weakness is that simulated physics differs from the real world, which is why randomizing conditions is needed.
Does 97 percent success mean RT-1 is reliable in any kitchen?
No. The figure applies to the instructions in its trained set under the paper's test conditions. The 76 percent result on unseen instructions is a more realistic sign of the gap in new situations.
Can one trained model now run any robot?
Not yet. Open X-Embodiment reported positive transfer between robots, but the claim is limited to the robots and tasks in its dataset.
Are these projects products readers can buy?
They are research results. Published papers describe experiments, and none of the figures here is a guarantee about a commercial machine.
Related reading
- How robots plan routes and navigate
- Computer vision and what robots see
- Deep learning behind robot skills
- Robots and AI in society
More in AI Basics.
Last verified: October 2026
Sources
- Solving Rubik's Cube with a Robot Hand, OpenAI et al., arXiv, 2019
- RT-1: Robotics Transformer for Real-World Control at Scale, Brohan et al., arXiv, 2022
- Open X-Embodiment: Robotic Learning Datasets and RT-X Models, Open X-Embodiment Collaboration, arXiv 2310.08864, 2023