The following article is from®÷ Intelligence Emergence, author Yang X✔&©iao

Since 2023, embodied intell★>£∑igence has been a hot topic that has at®₽tracted much attention.

In the view of Sun Teng, CEO of Ruoyu¶ £ Technology, robots are the most ♣λ→‍ideal landing carriers a→≈π nd application platfo¶  ↕rms for large models. The in÷$↑troduction of large mo"§÷ dels has completely c↓↕hanged the dilemma that robots caα&≤∏n only perform a single task in a si​‌ngle scenario, significantly improving ☆≈Ωthe intelligence level of r→↕βπobots, making them smarter and able to σ♥ undertake more complex tasks.

At present, some robot∞≥÷§ companies choose to load exte ÷ rnal large models (such as calling←  Chatgpt) directly on th↓→e robot. However, these large models a‍'≈re not trained with specific ☆✔δrobot-related data for rob ∞<ot embodied intelligence, which maβε≈y cause the large model to '↔☆be incompatible with the robo∏™→αt execution code and the ex '♦ternal environment, thereby reducing>'™ the instruction decompo↓✘&sition ability and e'★xecution success rate. At the same §  time, large models loaded in←§< the cloud will also bring abouπ​"♠t latency problems, restricting the d≠>✔σomain adaptability of multi-type ₹§ε±robots driven by large models.

Ruoyu Technology is committed to the rβ≤ esearch of universal robot brains, and®± hopes to cooperate with robot manufact≈ ≈urers such as indust£±₹rial robotic arms and service ro™>↕>bots to provide advaσ∞ε®nced robot brain products.

Ruoyu Technology' ∑φ s research on the general robotφφ± brain based on mult∞↑€↓imodal large models maiα₩nly focuses on two directions: one iλ₽>÷s the perception model ₩∞responsible for the perception le✔±₽‍vel, and the other is the con®Ω>δtrol model at the con<λ≠₩trol level. The perception mode₹γl obtains external environm ♥ent information through voicπεe interaction, visual infor×‌mation sensor signal extraction, et<®c.; after perceiving the•≈ environment, the control model conver£"δ©ts the received demands into♦λ control instructions that th•☆e robot can operate and execu ×φte, such as instruction di€÷γ$sassembly, code generation and otheΩδ₹•r control-level tuning work.

Ruoyu Technology and Harbin Institute oσ∞♣•f Technology jointly devγ∞λeloped the language large model ®φ↕ base - Lizhi and the multimoλ Ω→dal large model base - Jiutian,→δ♣ with a total parameter≠←→ scale of 13 billion and training data ‌®exceeding 1500B tokens. It can complet< ≈e multiple types of instruction≥✘ data such as multi-task, multi-round ↕©dialogue, Chinese-Engli÷₩sh translation, thinking ​ chain, tool use, etc., ₽ ¶and can show good resul ♥ts in logical reasoning, relational σγreasoning, and perception ability.

Based on the self-developed language-∞α​‍based big model and α↔‍multimodal big model, §  Ruoyu Technology has built a mul>✔timodal embodied decision-making b→♦ig model. With the univers​¥☆al knowledge of the h✔¥uman world and the powerful∏Ω← proprietary code generation capabili ↕→γties in the field of robotics, it can aσΩchieve a deep underst§♠ε✔anding of the extern​∞¥al environment and intelligent contro↕∑l of various types of robots. In₽γ♥ terms of product form, Ruoyu Techno₹© logy provides two typ'→✔es of standardized p∑₩roducts for robot manufact☆™urers: one is the sof£γ$tware form of SDK, which &♠is loaded on the public machin♦™e on the robot side or deployeφ<d locally; the other is to co"β↕mbine the robot brain wi<₩≥'th the computing chip to ₩₩αprovide a hardware and softwa ←re integrated board s©Ωolution.

The robot brain develo♥↑ped by Ruoyu Technology can be dir♠γ₹ectly deployed at the internal devel✔₩₽‍opment end of the robot arch☆ε¶αitecture, and communicate α‍directly with the robot's app' lication layer and middleware. By auto​ matically generating code, it >♥₹reduces the development workload™∑ of traditional robot ma∑φ'nufacturers and simplifies the develop ₹$$ment difficulty, which helps tradπ>​itional robot manufacturers Ω develop various different app≥∑↓lication functions.

△ Ruoyu Jiutian robot brain architect±€ure

Ruoyu Technology has chosen to coo§©perate with leading server man±∏ufacturers to cover 70% to>≥ 80% of scenarios in different fields t≤<δ♦hrough the core needs of server ma∞πnufacturers. In the futurε"e, the company plans > to further increase the functions"<←↓ required by customers when≤™☆≤ cooperating with mid- and long-tail c÷☆∑→ustomers.

Ruoyu Technology has completed the r βesearch and development of☆ε'€ a base model with 13 billion ₹ parameters, and has made‍≠♠ considerable progress in ↕>÷∏the design of the robo✔≤t brain architecture and produc≠₩>¥t design and development. The company <γ‍is developing functions for robot​₹>&s covering different fields and d↑¥ifferent forms, and exp↑​‍ects to complete the deε™velopment of the robot b> ≠÷rain in the first half of 2024.
In addition, the company will open<€ a free robot brain SDK in the second γ↔quarter of 2024 to better underst₽↔±∏and market demand and promot ¶₽₽e iterative products.

At the hardware level, ✘α★ Ruoyu Technology will also st↑↓λrengthen the connection between differ‍÷ent forms and interfac±☆es in the robotics field, realizi←‌δ₹ng the transition from execεφπuting instructions at"§ the data interface to collaboraφ± tive work with different ha®<rdware manufacturers.

As for the founding team, the fo₹‌♣λunding team of Ruoyu Technology comes ≠γ✔≥from the School of Computer S≤¥¥cience and Technology of Harbin Insφ∑αΩtitute of Technology (Shenzhen). The >​✔ team has dozens of national higΩ∏♥h-level talents in mul♥​σtimedia computing, na¥∏€‍tural language proces₽☆✔&sing, embodied intelligence •δ☆and large model quantization compressi→☆≤on, and relies on Sh★ &enzhen Ha Shen Asset Management Co., L$★£td. to transform scie↔♥ ™ntific and technological achie←'Ωvements. Sun Teng, the company's co∑>≠-founder and CEO, is a doct ↕or in artificial intelligence ★≥σand has continuous entrepreneurial e'​™÷xperience.

Message from Profess☆≈βor Zhang Min, Chief Scieπ₩Ωntist of Ruoyu Techn↔©•¥ology:

In the development of large mo‍™ dels, the importance of data cannot b>§§e ignored. The distribution×£≠, quality and input order of dat♠σ↑∞a will affect performance. Secondlyα₹, he emphasized the import→"Ωance of personally participating i€✘βn the development of large mode♣πls, believing that only by trying÷∑" it yourself can you b♣≈≥etter grasp the original§£α↓ innovation. Therefore, he suggeste∏γ★d that every university ♥‍₽δteacher should persona↑★₽lly train a large model, even if it ¶¶is small in scale, wi™∑th only 100 million ¶©σparameters. Analogous to the principle‌→  of manufacturing an engine, even if o→ thers make it well, it is crucial to ↓≠‍personally experience the process ₩Ω.

Modified on December 27, 2023. If th♦ ere is any infringement, pleas∏ &e contact us to make modifications>γ¶.