What does a robotic arm designed for "human-computer interaction" look like?

This robotic arm focuses on being "safe" and "flexible" and is even lighter than a human arm. It is a two-arm robot from South Korea - Ambidex.

In 2019, we introduced its prototype, which was developed by NAVER LABS. Its parent company NAVER is known as the "Google of South Korea".

At that time, Ambidex only had two arms. When we met again after 5 years, it looked more like a human.

The biggest change is that with the head + trunk + waist, the range of motion and stability have been greatly improved.

Ambidex has also upgraded its haptic devices and can now complete more complex tasks, such as helping its "colleagues" wash their faces:

I am also a good hand in the kitchen. I will peel a sweet potato for you to eat:

If you have nothing to do, let’s have some entertainment and play table tennis:

▍Cable drive + new minimalist style design

The AMBIDEX arm uses a cable-driven mechanism that places all the heavy-duty actuators on the shoulder and body parts, which reduces the weight of the arm. A single robotic arm weighs only 2.6 kilograms, which is lighter than an adult male's arm.

A lighter arm represents safety and flexibility, allowing it to safely interact with people, such as hugging and shaking hands:

Despite its light weight, AMBIDEX can withstand a weight of 3kg and can run at a maximum speed of 5m/s.

Like the human arm, the Ambidex arm also has 7 joint degrees of freedom. Officials say its innovative mechanism can simultaneously enhance the force and strength of each joint, enabling the same level of control, performance and precision as industrial robots!

After adding new head, torso and waist designs, Ambidex looks cooler. However, NAVER LABS said that they did not want to sacrifice research efficiency for the sake of external image, so they adopted a "minimalist style" in appearance. The design uses black areas to hide sensors and bolted components, reducing visual complexity.

At the same time, considering the movement of the robot and the convenience of assembly, NAVER LABS minimized the cover configuration and focused on the details of the parting line so that the cover can be separated, retaining the wiring while reinterpreting it as a design element, during the display While maintaining the basic structure, it maximizes the overall visual balance and unity.

NAVER LABS has not yet announced any more technical highlights of the "Humanoid Ambidex", but the movements do seem smoother and more flexible.

▍Learn human “physical intelligence”

Look carefully, when Ambidex performs various tasks, there is a staff member behind him who hides his fame and achievements:

It seems that the staff is controlling Ambidex, but in fact Ambidex is learning human "physical intelligence". Unlike most robotic arms that move based on pre-input position information, Ambidex is designed for "human-computer interaction" and must ensure absolute safety in unstructured daily tasks.

So what is human physical intelligence? NAVER LABS said: In addition to conscious activities when humans interact with the external environment, there are many things that are unconscious. For example: when smearing jam, humans do not consider every movement of the fingers or the tiny force exerted on the wrist and fingers. The knowledge and intelligence based on this physical experience are called "physical intelligence".

In order to impart "physical intelligence" to robots, NAVER LABS developed a tactile device. Unlike common control devices, Ambidex's tactile device provides position and force data in a bilateral manner. Not only does it allow the user to send commands to the robot to move its position, but it also enables the user to feel the forces the robot is experiencing. People control robots through tactile devices to perform tasks, and at the same time, robots indirectly acquire human physical intelligence.

▍Tactile equipment upgrades, from peeling sweet potatoes to playing table tennis

Let’s see how Ambidex uses tactile devices to peel sweet potatoes:

Although it seems simple, the entire action is performed by the force control of the robot and the bilateral force transmission of the haptic device. If either one doesn't work properly, excessive friction is exerted between the knife and the potato, potentially damaging the ingredient, the tool, or the robot itself.

“Peeling Sweet Potatoes” shows only Ambidex’s first-generation tactile device, which was subsequently upgraded by NAVER LABS. Haptic Device 2.0 has a larger movement space, and task performance for complex and detailed movements has also been significantly improved, including: the speed with which the robot responds to the operator's movements, and the efficiency with which users feel the forces felt by the robot.

The upgraded haptic device can perform complex tasks involving fast movements, such as playing table tennis, where the racket must be quickly positioned in the correct position to align with the incoming ball.

At the moment of impact (when the ball is struck), it must be moving with sufficient speed and in the correct direction. If the inertia of the haptic device was slightly greater, or if there was a delay in transferring the magnitude and direction of the force from the human movement to the robot, the ping pong demonstration would not be successful.

Although Ambidex has not yet achieved autonomous execution of tasks, perhaps this day is not too far away, because the goal of NAVER LABS is to use bilateral haptic devices as an effective tool to collect high-quality training data to improve the intelligence level of Ambidex and ultimately achieve autonomous execution. Task.

▍Conclusion

NAVER LABS’s technology roadmap for Ambidex is clearly visible. Collecting high-quality human-computer interaction data through bilateral tactile devices to train the robot's intelligent model. This idea of "human-machine collaborative learning" can be said to be an important technological innovation.

From the perspective of technical implementation, the quality and richness of data collected by bilateral tactile devices will directly affect the performance of robot intelligent algorithms. NAVER LABS needs to be further optimized in hardware design, sensor selection, signal processing and other aspects to ensure the accuracy, real-time and diversity of data. At the same time, how to design an efficient machine learning framework to extract common behavioral patterns and decision-making strategies from massive and complex interactive data is also a key challenge. This requires systematic research in the fields of deep learning, reinforcement learning, transfer learning and other fields.

In addition, data-driven intelligence improvement is closely coupled with the robot's electromechanical control system. If Ambidex wants to finally achieve autonomous operation, it needs to make new breakthroughs in motion planning, force control, and dexterous operation. For example, how to achieve precise grasping and operation based on visual and tactile feedback? How to maintain balance and robustness in a dynamic environment? These are technical difficulties that need to be overcome urgently.

In general, the technical blueprint developed by NAVER LABS for Ambidex is practical and feasible, and its development idea of "people leading machines and being data-driven" represents the cutting-edge direction of robotics research. Currently, Ambidex is still in the stage of technology accumulation and iterative optimization. To truly grow into a fully autonomous intelligent system, it needs to continue to make efforts at multiple levels such as perception, cognition, decision-making, and control.

humanoid Robot;

AI Robot;

AI service Robot.