Published on: March 18, 2025 / update from: March 18, 2025 - Author: Konrad Wolfenstein
Humanoid Standing-up Control: Learning to get up with hosts-the breakthrough for robots in everyday life-Image: Humanoid-Standup.github.io
More than just getting up: Host paves the way for autonomous and versatile humanoid robots
From the simulation to reality: How Host Humanoid robots teaches the self -employed up
In the fascinating world of humanoid robotics, in which machines imitate more and more human abilities, an apparently simple but fundamentally important skill is playing a central role: getting up. It is a matter of course for us humans, an unconscious movement that we perform countless times every day. But for a humanoid robot, getting up is a complex challenge that requires the interaction of sophisticated control, precise sensors and intelligent algorithms. However, this ability is not only an impressive demonstration of engineering art, but also an essential prerequisite for humanoid robots find their place in our everyday life and can support us in a variety of areas of responsibility.
Getting up from different positions is much more than just a nice additional function. It is the foundation for autonomy and versatility of humanoid robots. Imagine that a robot should help you in the household, assist in care or work in dangerous environments. In all of these scenarios, the ability to set up independently from different locations is of crucial importance. A robot that only works in ideal starting positions and remains helpless when falling is simply unusable in the real world. The development of robust and versatile up -to -up strategies is therefore a key step to bring humanoid robots from the research laboratory to the real world.
Previous approaches to solve this problem often reached their limits. Many were based on laboriously preprogrammed movements that worked in controlled environments, but quickly reached their limits in unpredictable reality. These rigid systems were inflexible, could not adapt to changed conditions and fail miserably when the robot landed in an unexpected position or was on uneven surfaces. Other approaches rely on complex simulation environments, the results of which were often difficult to transfer to real robots. The leap from the simulation to reality, the so-called “Sim-to-Real Transfer”, turned out to be the stumbling block of many promising research approaches.
In this context, an innovative framework enters the stage that could fundamentally change the way we think about getting up humanoid robots: Host, short for Humanoid standing-up control. Host is more than just another method; It is a paradigm shift. Developed by a consortium of renowned universities in Asia , including the Shanghai Jiao Tong University, the University of Hong Kong, Zhejiang University and the Chinese University of Hong Kong, Host breaks with traditional approaches and takes a completely new way to teach humanoid robots - in a way that is astonishingly versatile, robust and realistic.
Suitable for:
Host: A framework that learns from mistakes
The core of the host innovation lies in the use of Reinforcement Learning (RL), a method of machine learning that is inspired by the way people and animals learn. Imagine you teach a child cycling. They do not give him detailed instructions for every muscle movement, but simply let it try it. If the child falls there, it corrects his movements on the next attempt. Through attempt and errors, the child gradually learns to master the bike through positive and negative feedback. Reinforcement Learning works according to a similar principle.
In the case of Host, a humanoid robot is placed in a simulated environment and confronted with the task of getting up from different positions. The robot acts as a “agent” in this area. It performs actions, in this case movements of his joints and his body. For each campaign, he receives a “reward” or “punishment”, depending on how successful it was. If he gets up, he receives a positive reward. If it falls or makes unwanted movements, he receives a negative reward. Through countless attempts to gain experience and the optimization of its strategies, the robot gradually learns to develop the best possible stand -up strategy.
The decisive difference to previous RL-based approaches is that Host learns from scratch. No preprogrammed movements, no human demonstrations or other previous knowledge are used. The robot begins with an “empty sheet” and develops its up -to -date strategies completely independently. This is a fundamental progress, because it enables the system to find solutions that may go far beyond what human engineers could have come up with. In addition, the system makes it extremely adaptable because it does not rely on rigid assumptions or human bias.
The magic of the multi-critic architecture
Another heart of host innovation is the multi-critic architecture. To understand that, we have to briefly deal with the functioning of Reinforcement Learning. There are two central components in typical RL systems: the actuator and the critic. The actuator is, so to speak, the robot's brain that selects the actions, i.e. decides which movements should be carried out. The critic evaluates the actuator's actions and gives him feedback. He tells the actuator whether his actions were good or bad and how they can be improved. In traditional RL approaches there is usually only one critic.
Host breaks with this convention and instead relies on several specialized critics. Imagine there are different aspects when getting up that are important: hold balance, take the right posture, coordinate joints, control the rotating impulse. Each of these aspects could be evaluated by its own “expert”. This is exactly what makes the multi-critic architecture. HOST uses several critics networks, each of which specializes in a certain aspect of the starting process. One critic could, for example, rate the balance, another the joint coordination and a third party to the rotary impulse.
This division into specialized critics has proven to be extremely effective. It solves a problem that often occurs in traditional RL systems: the negative interference. If a single critic tries to evaluate all aspects of a complex task at the same time, conflicts and confusion can occur. The various learning objectives can hinder each other and slow down the learning process or even make it fail. The multi-critic architecture bypasses this problem by disassembling the learning task into smaller, clearer subtasks and using a specialized critic for each partial task. The actuator then receives feedback from all critics and learns to optimally combine the various aspects of getting up.
This multi-critic architecture is particularly relevant for the complex task of getting up. Getting up requires a variety of fine motor skills and precise control of the rotary impulse in order to keep the balance and not to fall over. Through the specialized critics, Host can specifically train and optimize these different aspects of getting up, which leads to significantly better results than conventional approaches with a single critic. In their studies, the researchers have shown that the multi-critic architecture enables a significant leap in performance and enabled Host to develop stand-up strategies that would be unreachable using conventional methods.
Curriculum learning: From the simple to the complex
Another key to Host's success is the curriculum -based training. This method is based on the human learning process, in which we gradually learn complex skills, starting with simple basics and then slowly working up to us. Think about the example of cycling. Before a child learns to drive on two wheels, it may learn to keep your balance on a impeller or drive with support bikes. These preparatory exercises make the later learning process easier and ensure faster and more successful progress.
Host implemented a similar principle. The robot is not confronted with the most difficult task right from the start, namely to get up on any surface from any position. Instead, it is subjected to a staggered curriculum in which the tasks gradually become more complex. The training begins with simple scenarios, for example getting up from a lying position on the flat floor. As soon as the robot has mastered this task well, the conditions gradually become more difficult. There are new starting positions on how to get up from a seated position or from lying on a wall. The surface is also varied, from level soil to slightly uneven surfaces to more demanding terrain.
This curriculum -based training has several advantages. On the one hand, it enables more efficient exploration of the solution space. The robot initially focuses on the basic aspects of getting up and learns to master them in simple scenarios. This speeds up the learning process and the robot reaches a good level of performance faster. On the other hand, the curriculum improves the generalization of the model. By gradually confronting the robot with more varied and complex tasks, he learns to adapt to different situations and to develop robust up -to -up strategies that work not only in ideals but also in real environments. The variety of training conditions is crucial for the robustness of the system in the real world, where unpredictable surfaces and starting positions are the rule and not the exception.
Suitable for:
Reality through movement restrictions
Another important aspect of host is taking into account real applicability. Simulations are a powerful tool for training robots, but the real world is unequal more complex and unpredictable. In order to successfully master the leap from the simulation to reality, Host implements two significant restrictions on movement that ensure that the strategies learned can also be implemented on real hardware and do not damage the robot.
The first restriction is smoothness regularization. This aims to reduce oscillating movements. In simulations, robots can carry out movements that would be problematic in reality. For example, they could make jerky, trembling movements that could be harmful to the physical hardware or would lead to unstable behavior. The smoothness regularization ensures that the learned movements are smoother and fluid, which is not only gentler for the hardware, but also leads to a more natural and stable stand-up behavior.
The second restriction is the implicit movement speed limit. This prevents too fast or abrupt movements. Here, too, simulations often represent idealized conditions in which robots could perform movements with unrealistically high speeds. In the real world, however, such abrupt movements can lead to damage to the robot, for example to overload the engines or damage to the joints. The movement speed limit ensures that the movements learned remain within the physical limits of the real hardware and do not endanger the robot.
These restrictions on movement are crucial for the SIM-to-Real transfer. They ensure that the strategies learned in the simulation not only work theoretically, but can also be practically implemented on real robots without overloading or damaging the hardware. They are an important step to bridge the gap between simulation and reality and prepare humanoid robots for use in the real world.
The practical test: Host on the Unitree G1
The real test for every robot control method is the practical implementation on real hardware. In order to demonstrate the performance of host, the researchers transferred the control strategies learned in the simulation to the UNITREE G1 Humanoid robot. The UNTREE G1 is an advanced humanoid platform that is characterized by its agility, robustness and realistic construction. It is an ideal test bed to evaluate the skills of host in the real world.
The results of the practical tests were impressive and confirmed the effectiveness of the host approach. The UNTREE G1 robot, controlled by host, showed remarkable impact capabilities from a wide variety of positions. He was able to successfully get up from a lying position, from a seated position, from the knees and even from positions in which he was leaning against objects or was on the uneven surface. The transmission of the simulated skills to the real world was almost smooth, which underlines the high quality of the Sim-to-Real Transfer from Host.
Particularly noteworthy is the robustness of disorders that the host-controlled Unitree G1 demonstrated. In experimental tests, the robot was confronted with external forces, for example by bumps or blows. He was confronted with obstacles that blocked his up. It was even loaded with heavy loads (up to 12 kg) to test its stability and load -bearing capacity. In all of these situations, the robot showed a remarkable resistance and was able to successfully set up without losing or overthrowing the balance.
In an impressive demonstration video, the robustness of Host became particularly clear. There you could see how a person bumped into the Unitree G1 robot during the starting process. Despite these massive disorders, the robot could not be removed. He corrected his movements in real time, adapted the unexpected effects and finally got up safely and stable. This demonstration impressively illustrates the practical applicability and reliability of the host system in real, unpredictable environments.
Suitable for:
Ablation studies: The interaction of the components
In order to examine the importance of the individual components of hosts more precisely, the researchers carried out extensive ablation studies. In these studies, individual elements of the host frameworks were removed or changed in order to analyze their influence on the overall performance. The results of these studies provided valuable insights into the functioning of hosts and confirmed the importance of the central innovations.
A central result of the ablation studies was confirming the decisive role of the multi-critic architecture. When the researchers modified the system in such a way that it only used a single critic, the system failed pitifully. It was no longer able to learn successful risks and the robot remained helpless in most cases. This result underlines the central importance of the multi-critic architecture for the performance of host and confirms that the specialized critics actually make a significant contribution to learning success.
The curriculum -based training also proved to be an important success factor in the ablation studies. When the researchers replaced the curriculum by random training without gradual increase in difficulty, the performance of the system deteriorated. The robot learned more slowly, reached a lower level of performance and was less robust compared to various starting positions and substrates. This confirms the assumption that the curriculum -based training improves the efficiency of the learning process and increases the generalization of the model.
The implemented movement restrictions also contributed significantly to the total output, especially with regard to practical applicability. When the researchers removed the smoothness regularization and the movement speed limit, the robot still learned in the simulation, but in reality they were less stable and led more often to fall or lead to undesirable, jerky movements. This shows that the restrictions on movement slightly restrict the flexibility of the system in the simulation, but are essential in the real world to ensure robust, safe and hardware -friendly behavior.
Host: A springboard for versatile humanoid robots
The ability to get up from different positions may seem trivial at first glance, but is actually a fundamental piece of puzzle for the development of really versatile and autonomous humanoid robots. It is the basis for integration into more complex locomotion and manipulation systems and opens up a variety of new applications. Imagine that a robot can not only get up, but also move seamlessly between different tasks - get up from the sofa, go to the table, grab objects, avoid obstacles and get up when he stumbles. This type of seamless interaction with the environment, which is a matter of course for us humans, is the goal of the humanoid robotics and host brings us a decisive step closer to this goal.
Host could be used with Host in the future in a variety of areas in which their human form and their ability to interact with the human environment are advantageous. In nursing, they could support older or sick people, help them get up and sit down, enough objects or assist in the household. In the service area, they could be used in hotels, restaurants or shops to operate customers, transport goods or provide information. In dangerous environments, such as disaster reliefs or in industrial plants, they could take on tasks that are too risky or too exhausting for people.
In addition, the ability to get up is also essential for stubborn production. Falls are a common problem with humanoid robots, especially in uneven or dynamic environments. A robot that cannot get up independently after a fall is quickly helpless in such environments. HOST offers a solution here because it enables the robot to reappear from unexpected locations and continue its task. This increases the reliability and security of humanoid robots and makes them more robust and more practical tools.
Host paves the way for a new generation of humanoid robots
Host is more than just a further development of existing methods; It is a significant breakthrough in the control of humanoid robots. Through the innovative use of Reinforcement Learning with multi-critic architecture and curriculum-based training, it overcomes the restrictions of previous approaches and enables robots to stand up from a remarkable variety of positions and on a wide variety of surfaces. The successful transfer from the simulation to real robot, demonstrates on the Unitre G1, and impressive robustness to disorders underline the enormous potential of this method for practical applications.
Host is an important step on the way to humanoid robots that not only impress in the laboratory, but can also offer real added value in the real world. It brings us closer to the vision of a future in which humanoid robots are seamlessly integrated into our everyday life, support us in diverse tasks and make our lives more comfortable, more comfortable and efficient. With technologies like Host, the once futuristic idea of humanoid robots that accompany us in our daily life becomes more and more tangible reality.
Suitable for:
Your global marketing and business development partner
☑️ Our business language is English or German
☑️ NEW: Correspondence in your national language!
Konrad Wolfenstein
I would be happy to serve you and my team as a personal advisor.
You can contact me by filling out the contact form or simply call me on +49 89 89 674 804 (Munich) . My email address is: wolfenstein ∂ xpert.digital
I'm looking forward to our joint project.
