All operating layers are connected to AI, and Menteebot brings new ideas to the development of humanoid robots

Enthusiast Network reported (Text/Wu Zipeng) Recently, the startup Mentee Robocs demonstrated its first humanoid prototype Menteebot. Compared with Boston Dynamics' electric Atlas and Optimus, Menteebot is not well-known.

However, Mentee Robotics is still famous because it is another entrepreneurial project of Mobileye's core founder Amnon Shashua. The other two founders are Mobileye's Chief Technology Officer Shai Shalev-Shwartz and former Facebook Research Director Li Wolf.

The biggest feature of Menteebot is that it is said to be connected to AI at all operating levels. Note the difference here. It is not to build a robot's "brain" around the AI ​​model, but to use AI to empower all operating levels.

Menteebot is full of AI

Mentee Robotics said that Menteebot is a "personalized robot that can be guided" with balance and control capabilities similar to humans. As mentioned earlier, unlike previous humanoid robots, Menteebot is full of AI and equipped with rich AI models and AI, which allows Menteebot to benefit more from the development of multimodal AI models. At the same time, humans can use natural language to control it.

Through the demonstration, Menteebot's super AI capabilities are manifested in at least two aspects, one is a more natural communication ability, and the other is super movement ability.

Let's look at the communication ability first. Menteebot is equipped with a richer natural language model, which makes its control no longer limited to established instructions. In the past, when humanoid robots communicated, there was a gap between the capabilities inside and outside the model. We have also mentioned this issue before. When a certain instruction is within the scope of model training, the humanoid robot can show super capabilities, but when the instruction exceeds the scope of model training, the humanoid robot will appear "at a loss".

It is reported that Menteebot can have natural conversations and exchanges with humans. Users only need to give commands to the robot through natural language, and it can understand and perform the corresponding tasks. This is actually due to the more powerful artificial intelligence algorithms, large language models and software in the Menteebot "brain" model. Mentee Robotics said that based on Ne's real-time three-dimensional mapping and positioning, dynamic navigation in complex environments and other technologies, complex reasoning can be achieved to complete tasks and quickly learn new tasks.

Mentee Robotics specifically mentioned that the deployed Menteebot can be trained at a higher level to cope with complex tasks or scenarios. At this time, the software will continue to perform tasks until it is mastered, and then the robot can complete the task in the real world.

Looking at the athletic ability, Menteebot can complete more complex walking postures, such as running, walking sideways, and even turning; it can also perform very delicate operations. It can accurately hand tableware to humans, thanks to the full range of motion and precision of its arms and hands.

In order to enhance the mobility of Menteebot, Mentee Robotics integrated cutting-edge 2Real learning on this robot, which can achieve realistic gait and hand movements, have the same balance and control as humans, and adjust gait when lifting heavy objects.

Mentee Robotics said that the mass production version of the Menteebot robot is expected to be deployed in the first quarter of 2025, and will be powered by pure visual sensing, dedicated electric motors that support "unprecedented" flexibility, and fully integrated artificial intelligence. It is expected to be divided into two versions: home version and commercial version.

Menteebot brings inspiration to humanoid robot innovation

By connecting AI to all operating layers, the Menteebot robot has demonstrated powerful capabilities, including communication and movement. But this actually puts higher requirements on the deployment of models and algorithms, including the core in the Menteebot "brain" and the chips in other execution units.

For the core chip in the "brain", it must first be able to support the deployment of multi-modal AI large models, and it is necessary to leave deployment space for new modalities that will be integrated in the future. At present, the large models used in humanoid robots include image recognition modules, modules, text-to-speech modules, dialogue system modules, navigation modules, multi-modal system modules, and reinforcement learning modules. The core chip is the carrier of these modules. For the above modules, or modalities, they need to be able to support them. Even if they are summarized, the core chip needs to be able to fully support the four major capabilities of vision module, navigation module, language module, and decision-making module, which puts high demands on the operator richness of the core chip. After the release of Menteebot, reinforcement learning is expected to rise to the fifth basic module. By then, software capabilities will continue to increase, and hardware redundancy will become an important indicator.

The second is how the execution unit can enhance AI capabilities, which puts new requirements on components such as and. Take MCU as an example. An important development direction of this category is to be able to deploy and execute AI programs on MCU. In fact, the industry has long begun to try to deploy AI on the smallest possible platform, but there are many challenges in this process.

First of all, to deploy AI models on MCU, these models need to be converted into C/code, which requires very accurate quantization of the models, which can not only ensure the capabilities of the models, but also try to avoid floating-point operations. This also places high demands on the MCU compiler, because after quantization, the code needs to be deployed to the MCU and needs to go through the compiler.

Secondly, when the MCU runs the AI ​​model, it cannot occupy too much on-chip resources and cannot have too high power consumption. Low power consumption is crucial because there are many MCU devices in a system, and too high power consumption will affect the overall endurance of the system.

Some people may say that MCU AI is unnecessary, but it is very necessary to achieve instant AI effects.

Conclusion

The launch of Menteebot is a dark horse in the humanoid robot industry, just as shocking as the electric Atlas. Connecting AI to all operating layers is an advanced concept, and it also requires chips as core hardware to be able to cope better. Among them, high-performance computing chips need to support multi-modality, and reinforcement learning modules are expected to rise to core modules; MCU AI is very helpful for improving immediacy, but it is also very challenging.