Electronic Enthusiast Network reports (by Liang Haobin): Recently, Xiaopeng Motors CEO He Xiaopeng posted a video on Weibo, showcasing the "Iron Assistant" operating a mobile phone screen to book the Xiaopeng MONA M03. In the video, the robotic hand moves smoothly, with a precise structure and high degree of freedom. The process of swiping the screen and pressing is very smooth, and it can quickly correct the operation after accidentally triggering text selection during the order process, giving off a bit of "true AI" flavor.

He Xiaopeng stated that the robot has a high degree of flexibility and tactile ability, and in the future, it can handle very delicate tasks in industrial scenarios. This year, he will also keep everyone updated on Xiaopeng's new progress in AI and large-scale hardware. Indeed, this year has seen the emergence of many new hardware pieces in the field of robotic hands and robots. With the assistance of AI, there may be hope for the "iPhone moment" of robotic hands.

With the surge of robotic hardware, the five-fingered dexterous hand has attracted attention.

Advertisement

Xiaopeng actually demonstrated a bipedal intelligent robot, PX5, at last year's 1024 Technology Day. At that time, according to the official introduction, PX5 could achieve top-level bipedal walking and obstacle-crossing capabilities in the industry. Through self-developed high-performance joints, it achieved high-stability robotic walking capabilities, capable of completing more than 2 hours of indoor and outdoor walking and obstacle crossing. In the demonstration video, PX5 showcased high-difficulty projects such as playing soccer and riding a balance bike.

As part of the bipedal robot, Xiaopeng is also developing humanoid robotic hands and mechanical arms. Among them, the humanoid dexterous hand has 11 degrees of freedom in one hand, and two fingers can pick up 1KG objects, while the fingers also have tactile perception capabilities, with a self-weight of only 430g; the ultra-lightweight humanoid mechanical arm has 7 degrees of freedom, with positioning accuracy that can reach 0.05mm, a maximum load of 3KG per arm, a self-weight of 5KG, and a maximum end-line speed of 1m/s.

With the dexterous hand and humanoid mechanical arm working together, Xiaopeng demonstrated operations such as picking up a game controller, using an electric screwdriver, lifting a box, pinching a ballpoint pen, pouring water, pulling paper, and pinching a plastic ball, but the movements were still slightly stiff.

From the video effect, this year Xiaopeng has made a more noticeable progress in the dexterous hand. First, the movements are more smooth and natural, possibly due to the new actuator, and secondly, the degree of freedom of the hand may have also increased. At Xiaopeng's 2024 Technology Day in two months, Xiaopeng may officially showcase the new dexterous hand and announce more details.

At the 2024 World Artificial Intelligence Conference (WAIC 2024) in July, Tesla showcased its second-generation Optimus humanoid robot prototype in China for the first time. The second-generation Optimus was first unveiled in December last year, when Tesla released a video demonstrating the capabilities of the second-generation Optimus.Tesla claims that the second-generation Optimus has lost 10kg compared to the previous generation, while its walking speed has increased by 30%. It has added some passive joints in the feet, making the walking motion appear relatively natural, although the actual walking speed is still quite slow. The key upgrade of the second-generation Optimus lies in its hands.

The robotic hand of the second-generation Optimus has 11 degrees of freedom, and the video demonstration shows very flexible and smooth movements. All fingers are equipped with tactile sensors, enabling precise operations, such as pinching an egg between two fingers. In the image above, the right side displays the pressure detected at the tips of the fingers of both hands.

Of course, a robot with only hardware cannot be applied in practice. To demonstrate the AI capabilities of Optimus, Tesla previously released a new video showing Optimus sorting battery cells on the production line of a battery factory. Optimus uses an end-to-end neural network, driven by FSD hardware computation, as well as visual sensors, hand tactile and force sensors, etc., to complete the corresponding tasks.

Speaking of robotic hands, the "strongest on Earth" humanoid robot just released by the star startup Figure AI is even more noteworthy. Figure, established in 2022, has investors including OpenAI, NVIDIA, Microsoft, and Jeff Bezos, the founder of Amazon. In August, Figure released the humanoid robot Figure02, which features an exoskeleton design and uses a new type of composite material with a honeycomb-like compressible structure near the knee and elbow joints. The actuators have also been upgraded; the arm joint actuator model A2 has a torque of 50NM, with a maximum joint angle range of 148°, and the lower limb leg joint showcases two types of rotary joint actuators: the L1 model with a torque of 150NM used near the hip joint of the thigh, with a maximum joint angle range of 195°; the L4 model with a torque of 150NM used at the knee joint, with a maximum joint angle range of 135°.

The key fourth-generation robotic hand boasts an impressive 16 degrees of freedom, capable of bearing 25 kilograms, with each finger driven by an integrated sensor and motor unit. Additionally, there is innovation in the wrist, providing a similar range of motion to the human wrist, while encapsulating and guiding all power and sensing signal lines to the hand and fingers. The team also stated that their goal is to enable the robot to grasp all objects that humans can grasp.

Another domestic robotics company, Unitree, has taken a different approach to building robots. Unitree launched its first general-purpose humanoid robot H1 last year, which has a relatively simple structure and no robotic hands, mainly used for testing humanoid robot walking algorithms. Unitree's advantage lies in its self-developed motor capabilities; the H1 knee joint uses Unitree's self-developed M107 joint motor, providing a peak torque of 360N·m, while the hip and ankle joints have torques of 220N·m and 45N·m, respectively, and the arm joints are 75N·m. All joint units of the H1, including servo motors, reducers, and controllers, are independently developed and produced by Unitree. The then price of $99,000 shocked the big names in the domestic and international robotics fields, and later, the H1 from Unitree could be seen in the robotics algorithm research of domestic and foreign universities.

In less than a year, Unitree has launched the G1, which is smaller, lighter, and has more degrees of freedom. The G1 basic version has 23 degrees of freedom without dexterous hands, and the EDU version can have up to 43 degrees of freedom with dexterous hands. Each leg has 3 degrees of freedom at the hip joint, 1 at the knee joint, and 2 at the ankle joint; each arm has 3 degrees of freedom at the shoulder joint and 2 at the elbow joint, with an additional option for 2 degrees of freedom at the wrist joint.

The G1 EDU is equipped with the Dex3-1 three-finger force-controlled dexterous hand, which is different from the common five-finger biomimetic hand on humanoid robots. The Dex3-1 has 3 degrees of freedom for the thumb, 2 for the index finger, and 2 for the middle finger, with the option to add 2 wrist degrees of freedom. The Dex3-1 three-finger force-controlled dexterous hand uses a hybrid force-position control technology, enabling the G1 to simulate the precise manipulation ability of the human hand, and it can also be equipped with a multi-point tactile array, achieving multi-point touch, simulating the delicate tactile sense of the human body, and significantly enhancing the ability to provide real-time feedback on tactile information.Although the performance of G1 from the video demonstration is already impressive enough, what is even more astonishing is its starting price of 99,000 yuan. Of course, if it is to be used for secondary development and applications with higher performance requirements, the G1 EDU version with the NVIDIA Jetson Orin high-computing module would be needed, and the price of this product has not been disclosed.

What are the key components of a robotic dexterous hand?

To achieve a human-like flexible hand structure, a robotic dexterous hand requires a variety of precision components for actuation. For instance, specific types of servomechanisms such as hollow cup motors are needed to provide the precise positioning and speed control required by the fingers, with high accuracy and response speed depending on the motor's performance.

Currently, the main suppliers of hollow cup motors overseas are companies like Faulhaber, Portescap, Allied Motion, Maxon Motor, and Nidec. In recent years, domestic suppliers have seen an explosive growth, with companies like Topband, Zhaowei Electric, Dingzhi Technology, Mingzhi Electric, and Weichuang Electric showing strong capabilities and beginning to enter the humanoid robot market.

As a component used in conjunction with motors, reducers are employed to decrease the motor's speed and increase torque output, which is also one of the key components in the field of robotics. In robotic dexterous hands, different types such as precision planetary reducers, RV reducers, and harmonic reducers may be used to meet various precision and torque requirements.

To achieve tactile feedback for the dexterous hand, a variety of sensors are needed, including torque sensors for measuring the force and torque applied to the fingers; tactile sensors for perceiving the shape, hardness, and texture of the contacted object; position sensors for monitoring the position and posture of the fingers; and temperature sensors for monitoring changes in the working environment's temperature and the object's temperature.

Taking tactile sensors as an example, the main companies in the global market currently are Tekscan, Pressure Profile Systems, Sensor Products, Weiss Robotics, Syn Touch, and others. However, the industrial chain in the field of robotics is also changing, with more players entering the market and launching innovative products in recent years. For instance, Melexis has introduced the tactile sensor Tactaxis, and the domestic company Pacini Perception Technology has launched the PX-6AX multi-dimensional tactile sensor.

The PX-6AX flexible array can capture sub-millimeter-level minute deformation fields instantaneously and also provides the robot with 15 types of multi-dimensional information such as sliding, friction, texture, and temperature. The sampling frequency can reach up to 1000Hz, with the ability to measure forces as small as 0.01N and offering a measurement lifespan of over 3 million cycles.

In summary, it can be observed that since the beginning of this year, humanoid robots have started to pay more attention to the dexterous hand part. The hand, being the most important part of the human body in work, will inevitably need to achieve human simulation on the machine hand if humanoid robots are to replace humans for general work in the future. At this time, there may still be some room for technical direction choices in terms of the walking method, such as wheeled feet. Therefore, whether the robotic dexterous hand will become a key direction for the industry to tackle in the coming period and be the first to be put into practical application, we can look forward to it with anticipation.