1. 首页Home
  2. 具身智能研发服务Embodied Intelligence R&D

03

具身智能研发服务Embodied Intelligence R&D

Embodied Intelligence R&D

提供从本体、算法到计算平台的具身智能研发服务。当前行业普遍采用“大脑 + 小脑”的分层架构：大脑负责多模态感知、场景理解与任务规划，小脑负责实时运动控制与底层闭环。我们据此提供本体搭建、基于视觉语言模型（VLM）的强化学习训练，以及计算方案集成。Embodied-intelligence R&D from the body and algorithms through to the compute platform. The field commonly uses a “brain + cerebellum” stack: the brain handles multimodal perception, scene understanding, and task planning; the cerebellum handles real-time motion control and the low-level loop. On that basis we provide body integration, reinforcement learning with vision-language models (VLMs), and compute integration.

咨询邮箱：Inquiries: jisuanjikexue1501@163.com

实验室中的人形机器人与机械臂示意

本体搭建Body integration

根据应用场景，完成传感器、执行器、机械结构与控制系统的选型与集成，搭建可运行的机器人或具身设备本体，为后续数据采集、算法训练和真机部署提供硬件基础。Based on the application, we select and integrate sensors, actuators, mechanics, and the control system to build a working robot or embodied device, giving a hardware foundation for later data collection, algorithm training, and deployment on the real machine.

基于 VLM 的强化学习训练VLM-based reinforcement learning

结合视觉语言模型的场景理解与任务规划能力，在仿真与真机环境中进行强化学习训练。覆盖策略学习、域随机化，以及从仿真到真机的迁移，形成可在本体上运行的控制与操作能力。训练可按任务设计奖励与评测，并与上层语言、视觉指令衔接。We combine the scene understanding and task planning of vision-language models with reinforcement learning in simulation and on the real machine. This covers policy learning, domain randomization, and sim-to-real transfer, so the body can run control and manipulation skills. Training can define rewards and evaluation by task, and connect to higher-level language and visual instructions.

大脑 + 小脑计算方案Brain + cerebellum compute

计算方案采用大脑与小脑组合：以国产化芯片或 NVIDIA Jetson Thor 作为大脑，承担感知、推理与任务规划；以 AMD 平台作为小脑，承担实时运动控制。这样把高层智能与硬实时控制分开，兼顾模型推理能力与运动闭环的稳定性、确定性。The compute plan pairs a brain with a cerebellum: a domestic chip or NVIDIA Jetson Thor as the brain for perception, reasoning, and task planning; an AMD platform as the cerebellum for real-time motion control. This separates high-level intelligence from hard real-time control, balancing model inference with a stable, deterministic motion loop.

  • 本体搭建Body integration
  • VLM 强化学习训练VLM reinforcement learning
  • 国产化芯片 / NVIDIA Jetson Thor（大脑）Domestic chip / NVIDIA Jetson Thor (brain)
  • AMD（小脑）AMD (cerebellum)

联系我们Contact us

合作与咨询，欢迎来信。For partnerships and inquiries, please write to us.

办公地点：天津市西青区399号天津市大学软件学院Office: Tianjin University Software College, No. 399, Xiqing District, Tianjin