requestId:681e3392ae8486.97795574.
Huaqiu SMT
High and reliable one-stop PCBA intelligent manufacturer
Huaqiu Mall
Self-operated electronic components mall
PCB Layout
High multi-layer, high-density product design
Steel Network Manufacturing
Special focus on high-quality steel network manufacturing
BOM unit
One-stop procurement and processing plan for special research
Huaqiu DFM
One-key analysis of design hazards
Huaqiu certification
Certification detection is indisputable
1 Introduction: What is embodied intelligence?
Have you ever thought that AI is no longer just staying on the screen to answer questions and write codes, but is truly entering the real world, perceiving, deciding, and moving like a human being?
Imagine in a smart warehouse, the robot plans the way to take the package from the shelf and pack and ship it. Imagine the robot at home that can not only scrub the floor, but also help you organize the room, wash dishes, and open the windows to see through. Behind this series of measures is Embodied AI, which focuses on: allowing AI to conceive, possess “mobile intelligence”, and have the ability to understand the surrounding conditions, perceive human instructions, and complete continuous operation.
In detail, embodied intelligence is an interlude of artificial intelligence, robotics, and superstition. It is important to discuss how to make robots have perception, planning, decision planning and action talents similar to humans [1]. Differences between traditional pure disk computing intelligence (such as speaking molds or image recognition), embodied intelligence is a great “bodyJamaicans Sugardaddy material” and the perception and interaction of surrounding conditions, and applies physical entities to perceive andModel the surrounding conditions, and only by relying on the purpose and the actual situation can the planning and decision-making plan, and finally apply the physical activities to complete the duties [2], and pay attention to the ability of AI to fulfill its duties in the actual situation.
The use of embodied intelligence is extremely common, and it is found in the following areas:
Industry dynamics: the mechanical arms complete precise grasping, disassembly, welding and other functions, and improve the effectiveness of giving birth.
Family office: The robots complete sweeps, deliver goods, helping the elderly and other functions, and improve the quality of career things.
Medical help: Surgery robots and rehabilitation robots help doctors complete rehabilitation or patient rehabilitation.
Exploration and rescue: Independent robots enter the risk area to perform detection and rescue operations.
Teaching and Education: Teaching robots help teach and accompany the robots to provide emotional interaction.
From “brain-type AI” (such as ChatGPJamaicans EscortT, Copilot) to “hand-type AI” (such as smart robot arms and home robots), this is the necessary path for the growth of artificial intelligence. After all, truly intelligent AI should not just “understand”, but also be able to “do”. As hardware costs drop, algorithms increase and data gathering, embodied intelligence will become the focus driving force in the intelligent period.
Figure 1 Embodied intelligence can be used in robots in various shapes
Origin: https://arxiv.org/pdf/2407.06886.pdfJM Escorts
2. The challenge of embodied intelligent research and development
Although embodied intelligent remote development, embodied intelligent research and development still faces many challenges, so that robots can efficiently improve new skills are not easy. The embodied intelligence in reality is far more complicated than imagined. Especially in the classical meaning of controlling the mechanical arms, even if it is “opening a door”, it is said to be “three disasters” for R&D personnel: 1. Building a scene: Simulating the surrounding situation.Set up a physical scene in the world, and the physical properties and initial conditions of the door are discussed. 2design measures: how to change the position of the mechanical arm, how to grasp the door, and where to open the door. 3 Writing the practice code: write the award function, adjust the excess parameters, and stop a large number of practice adjustments through the process. Each of the following cycles is highly dependent on manual interference, with a long opening cycle and low effectiveness. What’s more, each new skill is like building a car from the head. For example, if you hope that the robotics will “close the window” or “removal of the cup”, you have to rewrite and simulate the surrounding conditions, set up the installation and configuration parameters from the head, and even re-made the training logic. In summary, there are three major mountains behind the development process of embodied intelligence:
High labor cost: Each new skill requires a special research team to invest several weeks or even months, and to touch and simulate design, measure planning, algorithm adjustment and other multiple areas.
Low versatility: It is difficult to reuse other responsibilities for the surrounding conditions, measures and awards for specific duties.
Poor expansion: When the recurrence of duty increases (such as multi-object collaboration from single grabbing), the difficulty of opening increases exponentially, and it is difficult to iterate quickly.
Therefore, many embodied intelligent research and discussions in reality can only focus on a few fixed duties, which is difficult to expand rapidly.
3 Apply large speech model to create a “intelligent skills natural device”
In recent years, the Large Language Model (LLM) has demonstrated its reactionary potential in multiple areas with its strong speech understanding, common reasoning and code innate talent. Facing the challenges mentioned in the previous session, we thought: Can we not be able to use the powerful talent of “natural speaking + general intelligence” to embodied intelligence? Can we have the intelligence that can apply AI itself, actively implement the process, and significantly reduce the cost and maintain the effectiveness?
To this end, we have referenced several excellent opening projects for LLM applications [4] and combined them with the intact intelligent opening process. Finally, we released: Mu Xi’s embodied intelligent simulation natural system, completing the end-to-end active transformation from writing to skill improvement.
In detail, our application has a strong sense of speaking and common sense, combining the characteristics of robotic arm functions in embodied intelligence, designs a series of high-quality reminder words PromJamaicans Sugardaddypt templates, allowing it to be independent or born with new skills.Cleverly, understand the needs of the duties, convert the natural speech description Jamaica Sugar into a fulfilling simulation, born with all the inherent tasks required for natural power, and complete the highly active development of the mechanical arm technique process.
In short, it is a bit like a “embodied magician” – if you tell it what it wants to accomplish, it can take the initiative to create a complete set of fulfillment plans, from scenes to measures, from award numbers to simulation surrounding conditions, all of which are in place in one step, and it has completely changed the paradigm of embodied intelligence skills.
This system can or may actively complete the following meanings:
Innate skilled meaning drawing Description of the meaning of the object, the related meaning of the object is based on the object of the object of nature and the JM Escorts The status and properties of the object of nature and the attributes of the object of nature and the surrounding MuJoCo Innate meaning operation steps According to the meaningful skills and object, differentiate the meaning into a process sequence that can be performed by the robot arm. Inherent self-mechanical measures sequence. Disassembly the retrieval measures into the basic process unit of the robot arm. Inherent Glory Function Code. Inherent Glory Function Code. Inherent Glory Function Code. Inherent Glory Function Code. Inherent Glory Function Code. Inherent Glory Function Code. Inherent Glory Function Instance. Inherent Glory Function Instance. In the following process, you will no longer need to write manually. The scene XML setting installation, setting of various settings installation parameters, and modulator code logic required for MuJoCo simulation platform. You only need to inform the LLM scene of what manipulated objects are there, and the system can actively operate the power of multiple robotic arms to operate different objects, enter the complete training setting of each function, and then use the MuJoCo engine to start the simulation, and the robotic arms wil