Dexterous Robot-Arm Movement at Its Best: OpenAI Introduces New Dactyl System

As a toddler, I was fascinated by the grownups (around me) especially when it came to the ease at which they performed cumbersome tasks. As years went by, I learned that this accomplishment was achieved through experience and training as showcased by OpenAI researchers with their new Dactyl system.

Basically, the Elon Musk founded OpenAI unveiled the human-like hand, called Shadow Dextrous Hand, to manipulate and perform basic tasks such as playing with the child’s block.


Here’s why this unveiling is creating fuzz around it.

Mind-boggling working of the dexterous system

To achieve this outstanding precision in movement, OpenAI integrated some innovative system. Dubbed ‘dactyl’, the smart system encompasses algorithms and codes currently used in video games such as DOTA 2.

The human-like functionality of Dactyl system enables the AI robot to work in real-time. Moreover, the research includes three cameras to monitor the movement of the robot. That’s not all; the artificial intelligence robot is linked to a computer to relay real-time information on the functioning of the fingertip.

Additionally, the team had the daunting task of accumulating tons of data, amounting to 100 years of experience, and convert into 50 hours of work. To achieve this, 6,144 CPUs and eight robust NVIDIA v100 GPUs were used to power this mind-boggling innovation.

Rigorous testing of the Dactyl system

With this innovative system, the researchers attempted to programme the robots to move a six-sided cube from one location to another. However, things were twisted when it comes to the simulation. You see, in order for the robot to be extraordinary, it must be able to work in different environs.


In this case, the researchers at OpenAI ’messed’ up with the simulation to realize the desired outcome. This entailed changing the gravity, color of the cube and virtual hand, and randomizing the size of the cube. Nonetheless, testing the robot with varying gravity was the major highlight.

According to Matthias Plappert, a project member of OpenAI explained the importance of manipulating the working environ for the robot. “Without this randomization, it would just drop the object all the time because it wasn’t used to it,” he said.

Thinking outside the box

Conventionally, the machine learning system thrives on reinforcement training. Here, the robots are tried and tested to check on ensuring that they work seamlessly. Nevertheless, this has had its limitation. The most profound one is the challenge in the real-time application.

It is on this basis that OpenAI opted to incorporate the five-neural network to improve the performance of robots.

The remarkable success of OpenAI

As mentioned earlier on, the dactyl system borrows its algorithm from the OpenAI Five. This attributed to the great performance which entails state of the art techniques such as Proximal Policy Optimization (PPO) often used in games.

According to OpenAI, the bots are subjected to self-play to learn and build the functionality. In other words, the bots play millions of games amounting to experience estimated at 180years per day. This further required a lot of power as the company used 128,000 CPU cores to achieve this.


Nevertheless, this deep learning research enabled the robots to defeat humans in a game of Dota 2.

Hit or miss

So what’s the perception and view of the Dactyl system? Well, given the myriad of bots out there, Dactyl is quite unique. This is because of its capability to undertake one task and still change the orientation. Amazing right!!

Others may be troubled by the tremendous speed at which the researchers are developing robots to assimilate human behavior. As much as this being a genuine concern, OpenAI researchers are optimistic that their output is a breakthrough in general AI.

All in all, it will be interesting to watch this space and the pace at which the robotic systems adapt human-like behavior.