news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Don’t just look at Figure 02, the latest achievements of domestic robots have been exposed, showing off their muscles!

Right in Yizhuang, Beijing, there is such a humanoid robot "gym".

The robots here are the first to realize training on a treadmill.Can run 6km in one hourThe kind of:

And this is just a warm-up. You still have to climb stairs after running:

When we got outside, we also had a dedicated lane for robots, which could not stop in the 30-degree heat:

From time to time, we have to try some complex terrain such as lawns and slopes:

It is the pure electric drive full-size humanoid robot that was first launched in China in April this year.Heavenly Craftsmanship,Depend onBeijing Embodied Intelligent Robotics Innovation CenterResearch and development.

When it was first released, Tiangong’s main feature was anthropomorphic running, but within just a few months, its capabilities improved a lot.

In addition to running on a treadmill and walking on complex outdoor terrain as shown above, Tiangong now alsoAdded a large model, can communicate in Chinese and English, and can also follow human instructions to grab objects.

Recently, the 2024 World Robot Conference is about to be held. Quantum位 has learned in advance that an evolved version of "Tiangong" integrating "more than" these capabilities will be unveiled at the conference.

And by then, Tiangong will change his "vest", install it, and upgrade it.Large joints。

We spoke to Yijie Guo, head of motion control, about the capabilities of the evolved Tiangong.

42 degrees of freedom throughout the body, large model support

The upgraded Tiangong function parameters are as follows:

Compared with before, as shown in the following table, the height is 163cm and the weight has increased from 43kg to 56kg.

The whole body's degrees of freedom have increased to 42Compared with the previous model, the degrees of freedom of a single arm increased from 3 to 7, and the neck also increased by 3 degrees of freedom.

△The first generation Tiangong function parameters

Previously, Tiangong had no hands, but now it has them. Each hand weighs 600g, has 6 degrees of freedom, the tactile sensor has an accuracy of less than 0.3N, and the gripping force of a single finger is greater than 1kg.

And Tiangong is also equipped withFour full-scene binocular structured light 3D cameras, high-precision six-dimensional force sensor, which can achieve 360-degree environmental perception.

By sensing the environment, Tiangong can adapt to complex terrain and move through grass, sand, hills, and gravel, and can cope with a terrain difference of 30cm without any problem.

The running stability is also improved, and the speed can reach 6km/h.

In addition, another major upgrade of Tiangong is that it is equipped with a large model, which hasMulti-modal operation capability。

It can now speak and interact with humans, supporting Chinese and English.

It can also follow human instructions to grab items:

[Unable to insert the video here, unfortunately... You can check it on the QuantumBit official account~]

And these are just some of them, there will be more displays and surprises at the World Robot Conference.

When Tiangong was first released, it focused on the single ability of anthropomorphic running. From the current evolutionary version, combined with sensory interaction, it can be said that Tiangong has initially formed aEmbodied Agentform.

Researching embodied intelligent planning, decision-making and task execution is also the current focus of the team behind Tiangong.

Building embodied intelligence

Based on embodied intelligence, the research team currently focuses on improving Tiangong’sVisual perception ability。

Before, Tiangong walked in a "blind" state and needed to use the soles of his feet to test the ground. But now, based on visual perception, when faced with large terrain differences, Tiangong can respond by predicting in advance and taking actions such as raising his legs.

In terms of specific methods, the team developed its own sports skill learning method based on reinforcement learning:Predictive reinforcement imitation learning based on state memory。

This method has been introduced by Quantum Bit before. It combines the advantages of high stability of traditional methods with the advantages of strong generalization and independence of the environment of reinforcement learning.

It not only solves the problem of poor positioning accuracy caused by reinforcement learning, but also solves the problem of poor adaptability to unstructured environments in model predictive control methods.

Tiangong motion control director Guo Yijie also revealed to QuantumBit that the team had previously discovered during training that some networks may be easily disturbed by sensor drift during actual operation, and sometimes exhibit unstable posture, which led to the proposal of this method.

After the first generation of Tiangong was released, the team added more features to Tiangong during the following months of training.Historical state memory, so that it can estimate its current state and environmental terrain, thus achieving better generalization effect.

Guo Yijie also said that the current problem that embodied intelligence needs to solve is "Action"Question planning and execution issues.

In terms of task types and complexity, it can cover most tasks in human daily work and life. The more types of tasks it can do, the more complex and long-term tasks it can accomplish with the cooperation of the brain.

He shared several technical routes for implementing complex task planning:

For example, Tesla robots mainly collect data and then use the data for supervised learning, training robots to automatically execute in relatively simple and fixed scenarios. Although this method is relatively effective, its generalization ability is relatively poor.

Another method is to conduct reinforcement learning training in a simulation environment, which mainly relies on continuous trial and error and self-learning in an environment. The main problem encountered by this method is how to migrate from the simulation environment to the real physical scene. From the perception level to the specific physical interaction, there are great differences between simulation and real scenes.

Another method is to directly use the large model to output some task points, and then use traditional motion planning to perform these tasks.

The Tiangong R&D team's solution is to combine different methods. The next step is to create aMeta-skill Library：

At this stage, I think we need to expand the robot's skill library... including these methods, each of which can be used to solve different task scenarios. So each skill in the skill library may be implemented in a different way.

Solving the "common" problems of domestic robots

Let’s talk about the company behind Tiangong, which Quantum位 has introduced before.

Beijing Embodied Intelligent Robotics Innovation Center(hereinafter referred to as the Innovation Center), formerly known as Beijing Humanoid Robot Innovation Center, was established in November last year and was jointly established by Xiaomi Robotics, UBTECH Robotics, Beijing Mechanical and Electrical, Yizhuang Robotics, etc.

They are aiming atSolving the key common problems of embodied intelligent robots, avoiding the domestic robotics industry from simply reinventing the wheel.

The Innovation Center has brought together a group of top scientists and engineers, and also took the lead in establishing the Innovation Center Expert Committee and the Beijing Humanoid Robot Industry Alliance.

The Innovation Center Expert Committee is chaired by Qiao Hong, an academician of the Chinese Academy of Sciences, and deputy chairmen are Zhu Songchun, president of the General Research Institute, Huang Tiejun, chairman of the Artificial Intelligence Research Institute, and Wei Ran, chief engineer of the China Academy of Information and Communications Technology.

In April this year, the Innovation Center launched"Tiangong" universal robot mother platform, that is, an open source platform for embodied intelligent hardware.

The "Tiangong" platform can realize flexible expansion of software, hardware and other functional modules, and provides a series of open interfaces. Scientific research institutions and robot-related companies can carry out secondary development based on the software and hardware functions of the "Tiangong" mother platform.

At that time, they announced that they would focus on exploring a universal embodied intelligence platform driven by a large model.Multi-functional embodied intelligent body platform "Kaiwu", has gradually surfaced and is under intensive development.

The “Kaiwu” platform focuses onEmbodied Intelligence Big Model and Framework, focusing on multimodal key methods, embodied intelligent simulation application construction and the construction of a complete tool chain system.

Around "Kaiwu", the team is buildingLarge-scale embodied intelligence datasetIt is used to support the training and tuning of a large multimodal model of vision and language with over 7 billion parameters, and to realize the robot's Chinese interaction, open question and answer, scene visual understanding, embodied operation and other capabilities.

The plan is to release it by the end of 20252 million high-quality data。

Having said that, there have been a lot of new developments in the field of humanoid robots recently.

UBTECH revealed that humanoid robots have been working in the Zeekr factory; Tesla's official photos show that Optimus has entered the factory to pick batteries; OpenAI-powered robot Figure 02 has entered BMW to work...

Although "Tiangong" is not on the same track as them, and its main focus is to solve common problems of embodied intelligence, but... it cannot escape the fate of moving bricks.

It is understood that it is initially conducting data collection trials in some scenarios, such as carrying and sorting items that are relatively repetitive and boring, and conducting inspections and search and rescue in dangerous scenarios such as mines and construction sites...

The 2024 World Robot Conference will be held soonAugust 21Opened in Beijing Economic and Technological Development Zone, those who are looking forward to the evolution of "Tiangong" can stop by. I heard that the Innovation Center will have a wonderful booth where you can see the collective appearance of the "Tiangong" robot family, and there will also be a sub-forum~

One More Thing

Why do we have robots running on treadmills? Take a guess (doge).

Answer: In addition to having higher requirements for balance, the treadmill has controllable speed and can perform relatively accurate and objective tests. Robot running experiments require a relatively large area, while treadmills save space.