Embodied AI Glossary中文07Software & Tooling
The software that ties body, control, and perception together: ROS 2, URDF, motion libraries, learning frameworks, and deployment tools. · 232 terms
- 7.1Development environment basics11
- 7.2Getting started with ROS 223
- 7.3Robot models and coordinate frames17
- 7.4Communication middleware, in depth23
- 7.5Connecting real hardware, real time23
- 7.6Kinematics, planning, and control libraries30
- 7.7Perception, mapping, and navigation17
- 7.8Data logging and visualization10
- 7.9Deep learning and robot-learning libraries29
- 7.10Training engineering and compute14
- 7.11Deployment and inference speedups21
- 7.12Industry platforms and ecosystems14
7.1Development environment basics
First set up your machine: programming languages, Linux, code repositories, environment isolation, GPUs, and remote access.
- Python and C++Python 与 C++（具身常用编程语言）The two languages embodied AI relies on most: Python for models and experiments, C++ for real-time control.
- Ubuntu LinuxUbuntuThe Linux distribution maintained by Canonical that serves as the default operating system for robotics development.
- Windows Subsystem for Linux 2 (WSL2)WSL2（Windows 下的 Linux 子系统）Microsoft's official way to run a full Linux environment inside Windows, without dual-booting.
- GitHubGitHub（代码托管平台）The world's largest code-hosting platform, where nearly all paper code and open-source robotics projects live.
- A widely used Python package and virtual-environment manager that isolates dependencies between projects.
- An extremely fast Python package and virtual-environment manager written in Rust.
- Package Mirror Sources in China (pip / conda / apt / Docker mirrors)国内镜像源（换源：清华 TUNA / 阿里云等）Redirecting pip, conda, apt, and similar tools to a Chinese mirror instead of the slow or blocked default servers.
- A cross-platform build-configuration tool for C/C++ projects that spells out exactly how to compile them.
- Packaging a program together with its runtime environment into a container that runs the same way on any machine.
- NVIDIA's general-purpose computing platform for GPUs, which is what lets deep learning run on a graphics card.
- Secure Shell (SSH)SSH 远程登录An encrypted network protocol for logging into and controlling another computer's command line remotely.
7.2Getting started with ROS 2
With your environment ready, learn robotics software’s ‘glue’: nodes, topics, and other communication patterns, plus how projects are organized.
- Robot Operating System机器人操作系统The most widely used open-source middleware framework for robot development — not actually an operating system.
- The ground-up rewrite of ROS that defaults to DDS for its underlying communication; the standard choice for new projects.
- ROS DistributionROS 发行版A versioned, fixed bundle of ROS core libraries and packages, with each release tied to a specific Ubuntu version.
- FishROS One-Click Installer鱼香 ROS 一键安装A one-line install script from the Chinese ROS community FishROS that sets up ROS and its environment.
- A single-purpose program unit in ROS; a whole system is made of many nodes talking to each other.
- A named channel for continuously streaming data between ROS nodes, working on a publish/subscribe model.
- Publish-Subscribe发布/订阅Senders publish messages on a topic and receivers subscribe to it, with neither side needing to know the other.
- The agreed-upon data format that ROS nodes use to exchange information.
- cmd_vel Topic (geometry_msgs/Twist velocity command)cmd_vel 速度指令话题（Twist 消息）ROS's de facto standard topic for sending a mobile robot linear and angular velocity commands.
- ROS's request-response communication pattern: a client sends a request and the server sends back a result.
- A ROS communication style for tasks that take a while, with progress feedback partway through and the option to cancel.
- ROS Parameter Server参数服务器A shared key-value store in ROS 1 that holds global configuration values any node can read or write.
- Launch File启动文件A ROS script that starts multiple nodes at once and configures their parameters.
- The programming API used to write ROS nodes — rclcpp for C++, rclpy for Python.
- The basic unit for organizing code in ROS — one package groups a set of related nodes and configuration.
- ROS WorkspaceROS 工作空间A directory convention for holding, building, and installing your own ROS packages.
- The recommended build tool for ROS 2, compiling an entire workspace with a single command.
- amentament（ROS 2 构建系统）ROS 2's package build system, with ament_cmake for C++ packages and ament_python for Python packages.
- catkincatkin（ROS 1 构建系统）ROS 1's official build system, using CMake to compile and manage packages in a workspace.
- A ROS command-line tool that installs a package's system dependencies automatically from its declarations.
- ROS Master (roscore)ROS 主节点The central process in ROS 1 that registers nodes and helps them find each other.
- A bridge program that lets ROS 1 and ROS 2 nodes exchange topics and services with each other.
- Guyuehome (Chinese ROS community)古月居（ROS 中文社区）An early, sizable Chinese-language learning community and tutorial site for ROS robotics.
7.3Robot models and coordinate frames
Once you know ROS, describe a robot with files like URDF, then visualize it with TF and RViz.
- Unified Robot Description Format统一机器人描述格式An XML file format that describes a robot's links, joints, geometry, and mass, used as a standard across ROS.
- Xacro (XML Macros)Xacro（XML 宏）An XML macro language that adds variables, macros, and includes to URDF, cutting down on repetitive robot descriptions.
- An XML file that adds planning groups and collision exceptions on top of a URDF.
- The XML model format the MuJoCo simulator uses to describe robots and scenes.
- Google DeepMind's curated collection of high-quality MuJoCo robot models.
- Simulation Description Format (SDFormat)SDFormat 仿真描述格式The XML format used by Gazebo that can describe a robot and an entire simulated world.
- Pixar's open-source 3D scene-description format, now the standard scene format for NVIDIA's simulation platforms.
- A 3D file that describes an object's shape as a collection of triangles — the visual “shell” of a robot model.
- CAD Software (SolidWorks / Onshape / Fusion 360)CAD 建模软件（SolidWorks / Onshape / Fusion 360）Software for drawing 3D models of mechanical parts and assemblies, the starting point for designing a robot’s structure.
- SolidWorks to URDF ExporterSolidWorks 转 URDF 插件A plugin that exports a SolidWorks assembly directly into a URDF robot model.
- Open-source tool that exports robot assemblies built in Onshape directly into URDF or MJCF files.
- Free, open-source 3D modeling and rendering software, often used to process robot models and generate synthetic data.
- A tool that automatically slices a mesh into a set of convex chunks for use as a physics engine's collision shape.
- TF / tf2 Transform TreeTF 坐标树ROS's tree structure for tracking and querying the live transform between every pair of coordinate frames.
- ROS package that reads a URDF and joint angles, then publishes each link's coordinate transform.
- REP 105REP 105 坐标系约定（map / odom / base_link）The ROS convention defining the map, odom, and base_link coordinate frames for mobile robots.
- RViz / RViz2RVizROS's built-in 3D visualization tool for viewing a robot model, coordinate frames, and sensor data.
7.4Communication middleware, in depth
Digging back into the communication layer: DDS, QoS, zero-copy, and communication libraries beyond ROS.
- Middleware中间件The software layer sitting between the operating system and applications that handles communication between programs.
- A publish/subscribe communication standard from the OMG, and the default underlying transport for ROS 2.
- The ROS 2 settings that control how a message is delivered — reliably or not, how many kept.
- ROS 2 Domain IDROS_DOMAIN_ID（域 ID）A ROS 2 environment variable that decides which nodes can discover and talk to each other.
- ROS Middleware Interface (RMW)ROS 中间件接口The ROS 2 abstraction layer that separates the upper-level API from the actual communication backend.
- eProsima Fast DDSFast DDSAn open-source DDS implementation from the Spanish company eProsima, the default middleware in several ROS 2 releases.
- Eclipse Cyclone DDSCyclone DDSThe Eclipse Foundation's open-source DDS implementation, one of the optional communication middlewares underlying ROS 2.
- Eclipse ZenohZenohAn open-source communication protocol unifying publish-subscribe, storage, and query, usable in ROS 2 as a DDS replacement.
- Passing large data between processes by sharing a memory location instead of copying the content.
- Eclipse iceoryxiceoryxA shared-memory, zero-copy inter-process communication library often used by DDS and ROS 2 to move large data.
- Loading several ROS 2 nodes into a single process, so they pass data as raw pointers instead of copying it.
- ROS 2 Executor & Callback GroupsROS 2 执行器与回调组The ROS 2 machinery that decides which threads run callbacks, and whether they can run concurrently.
- A ROS 2 node type with a standard state machine, letting an external manager configure, start, pause, and shut it down in order.
- A lightweight publish-subscribe messaging library over UDP multicast, low-latency and common in real-time robot control.
- A lightweight, open-source messaging library that lets programs exchange data in just a few lines.
- Protocol BuffersProtobufGoogle's open-source, language-neutral format for serializing structured data compactly and quickly.
- Google's open-source RPC framework, letting programs on different machines call each other like local functions.
- A protocol for sending messages both ways, in real time, over a single TCP connection.
- Web Real-Time Communication (WebRTC)WebRTC 实时音视频传输A browser-native standard for low-latency, peer-to-peer audio, video, and data transfer.
- A WebSocket-and-JSON bridge that exposes ROS to web pages and other non-ROS programs.
- Model Context Protocol (ROS MCP Server)MCP / ROS-MCP-ServerA bridge that lets large models call ROS robot interfaces through a single, unified protocol.
- A Rust-based robotics dataflow middleware positioned as a lower-latency alternative to ROS.
- AimRT (AgiBot robot runtime framework)AimRT（智元机器人运行时框架）AgiBot's open-source, modern C++ robot runtime framework, interoperable with ROS 2 and other communication backends.
7.5Connecting real hardware, real time
With communication working, connect real hardware: vendor SDKs, drivers, embedded and real-time systems that get commands to the motors.
- Software Development Kit软件开发工具包The libraries, interfaces, and examples a manufacturer provides so developers can control a device from code.
- Writing your own programs against a vendor's SDK or interface to customize a robot, without touching its firmware.
- Unitree SDK2宇树 SDKUnitree's official open-source development kit for reading sensors and commanding joints and motion on its robots.
- Franka's official low-level C++ control library, exchanging commands with the robot in real time at 1 kHz through the FCI.
- Universal Robots' real-time interface that lets external programs read arm state and send commands.
- Intel RealSense SDK 2.0 (librealsense)RealSense SDKThe official open-source driver and development library for RealSense depth cameras, handling capture, alignment, and point clouds.
- The layer of software that lets an operating system or higher-level program read and write to a piece of hardware.
- A layer of code that hides specific hardware differences and exposes a uniform interface to the software above it.
- The standard ROS 2 control framework that connects control algorithms to real motor hardware.
- Writing the program that runs on the chip inside a device such as a motor driver board or a sensor.
- Compiling on an x86 computer a program meant to run on a robot's ARM-based onboard controller.
- Firmware固件The program burned into a hardware chip that directly controls a device's lowest-level behavior.
- Updating a device's firmware, system, or models remotely over a network, without opening it up or plugging in a cable.
- An operating system that guarantees tasks finish within a set deadline, used for timing-sensitive control.
- A guarantee that every single execution finishes before its deadline — missing even once counts as a system failure.
- A small open-source real-time operating system kernel that runs on microcontrollers.
- An open-source embedded real-time operating system developed by a team in China.
- A framework that brings ROS 2 onto microcontrollers.
- PREEMPT_RT实时内核补丁A real-time patch set that makes the Linux kernel almost fully preemptible, cutting worst-case latency.
- An open-source framework that adds hard real-time capability to Linux, common in low-level robot control.
- EtherCAT Master (SOEM / IgH)EtherCAT 主站The end of an EtherCAT bus that issues commands and gathers data, usually running on the main control computer.
- The Linux kernel's built-in CAN bus driver, which makes a CAN port behave like a network interface.
- Precision Time Protocol精确时间协议A synchronization protocol that aligns multiple devices' clocks over Ethernet to sub-microsecond precision.
7.6Kinematics, planning, and control libraries
Once hardware can send and receive commands, use existing libraries for kinematics, collision-free planning, and optimal control.
- MathWorks' commercial numerical computing and graphical modeling-and-simulation software, widely used in controls.
- Robotics Toolbox for PythonRobotics Toolbox for Python（Peter Corke 机器人工具箱）Peter Corke's Python robotics toolbox, a popular way to learn kinematics and dynamics hands-on.
- The most widely used open-source C++ linear-algebra library, behind most matrix, vector, and rotation math in robotics code.
- An open-source rigid-body dynamics library for fast robot kinematics, dynamics, and their derivatives.
- Open-source C++ library that efficiently computes a robot's forward and inverse dynamics.
- Orocos KDLKDLA C++ kinematics and dynamics library from the Orocos project, long a default choice for kinematics in ROS.
- An open-source numerical inverse-kinematics solver that's more reliable and faster than KDL.
- OpenRAVE's analytical IK generator, auto-producing extremely fast C++ inverse-kinematics code for a specific arm.
- PinkPink（基于 Pinocchio 的微分逆运动学库）A Python differential inverse-kinematics library built on Pinocchio, solving joint velocities from weighted tasks.
- An open-source Python library for differential inverse kinematics, built on MuJoCo.
- UC Berkeley's open-source JAX toolkit for solving robot kinematics optimization problems.
- An open-source motion planning library from Rice University that bundles dozens of sampling-based planning algorithms.
- The most widely used open-source C++ collision-checking and distance-computation library in robot motion planning.
- The most widely used open-source motion planning framework for robot arms, built on ROS.
- Open-source library that takes a path and computes the fastest way to traverse it within limits.
- Open-source library that generates jerk-limited, time-optimal trajectories in real time, every control cycle.
- NVIDIA's GPU-parallel motion-planning library that produces collision-free arm trajectories in milliseconds.
- NVIDIA Isaac ROS cuMotioncuMotionNVIDIA's GPU-based ROS 2 motion-planning package for robot arms, built on cuRobo.
- A framework that combines a symbolic planner with continuous samplers to do task and motion planning.
- A model-based robotics simulation and optimization toolbox led by MIT and the Toyota Research Institute.
- An open-source symbolic toolkit for numerical optimization and automatic differentiation, widely used for MPC and trajectory-optimization modeling.
- COIN-OR's open-source large-scale nonlinear solver, based on the interior-point method, commonly used for trajectory optimization.
- Open-source quadratic-program solver based on operator splitting, widely used in MPC and whole-body control.
- Open-source quadratic-program solver based on an online active-set method, commonly used in MPC.
- An open-source solver for real-time optimal control and MPC that can generate embeddable C code.
- A C++ library that solves whole-body control as a task-space inverse-dynamics quadratic program.
- A contact-aware optimal control library that solves robot trajectories using differential dynamic programming.
- Open-source C++ optimal-control toolbox from ETH Zurich, commonly used for MPC on legged robots.
- legged_control (NMPC + WBC framework for legged robots)legged_controlAn open-source legged-robot control framework combining nonlinear MPC planning with whole-body control tracking, built on OCS2 and ROS.
- MuJoCo MPC (MJPC)MuJoCo MPCGoogle DeepMind's open-source tool for interactive, real-time model predictive control inside MuJoCo.
7.7Perception, mapping, and navigation
Helping the robot understand its environment: image and point-cloud libraries, calibration, SLAM mapping, and autonomous navigation.
- The most widely used open-source computer vision library, for image, camera, and geometry processing.
- An open-source library for processing point clouds, meshes, and 3D reconstruction.
- An open-source C++ library for processing 3D point clouds — filtering, registration, and segmentation included.
- ETH Zurich's open-source calibration toolbox specializing in multi-camera and camera-IMU intrinsic, extrinsic, and time calibration.
- An open-source 3D reconstruction pipeline combining structure-from-motion and multi-view stereo.
- Nerfstudio / gsplatNerfstudio / gsplat（NeRF 与高斯泼溅开源工具库）Open-source toolkits for reconstructing 3D scenes from photos, covering both NeRF and Gaussian splatting.
- Google's open-source C++ library for nonlinear least-squares optimization, a common backend for SLAM and calibration.
- g2o (General Graph Optimization)g2o 图优化库A C++ library for solving graph-structured optimization problems like SLAM pose graphs and bundle adjustment.
- A factor-graph optimization library from Georgia Tech, a common backend for SLAM and sensor fusion.
- Common ROS SLAM Packages (GMapping / SLAM Toolbox / RTAB-Map)ROS 常用 SLAM 建图包（GMapping / SLAM Toolbox / RTAB-Map）Three ready-to-run ROS packages for mapping and localization that just need a config file, not custom code.
- Google CartographerCartographerGoogle's open-source real-time lidar SLAM system, supporting both 2D and 3D mapping.
- ROS 2's official navigation framework, letting a mobile robot autonomously drive itself from point A to point B.
- BehaviorTree.CPPBehaviorTree.CPP（C++ 行为树库）The most widely used C++ behavior-tree library in robotics, describing task logic in XML.
- Groot2 (BehaviorTree.CPP IDE)Groot2（行为树可视化编辑器，非英伟达 GR00T）A graphical editor and debugger for the behavior trees used by BehaviorTree.CPP.
- Autoware (open-source autonomous driving stack on ROS 2)Autoware（开源自动驾驶软件栈）A full open-source autonomous-driving software stack built on ROS 2, maintained by the Autoware Foundation.
- The two leading open-source autopilot software stacks, used on drones, fixed-wing aircraft, and uncrewed ground or water vehicles.
- Upper-level software that assigns tasks, plans routes, and prevents conflicts across a whole fleet of robots.
7.8Data logging and visualization
Recording and visualizing sensor and robot state: useful for debugging, and for collecting data to train on later.
- ROS BagrosbagROS's data-recording tool, which logs topic messages with timestamps so they can be replayed later.
- FFmpegFFmpeg（音视频编解码工具）An open-source audio and video processing toolkit, used for transcoding, cutting, frame extraction, and compression.
- ROS's built-in, plugin-based set of graphical debugging tools.
- Open-source time-series plotting tool commonly used to inspect ROS topics and log data.
- Foxglove StudioFoxgloveA robot data visualization tool that replays rosbags and shows curves, point clouds, and images in one view.
- An open-source visualization tool for multimodal, time-series robot data — images, point clouds, and curves played back in sync.
- A lightweight visualization tool that displays robots and 3D scenes in a web browser.
- A Python library for showing 3D scenes and interactive controls right in the browser.
- VuerVuer（网页 3D / XR 可视化与遥操作工具）A Python-driven 3D visualization framework that runs inside a VR headset's browser.
- QtQt（上位机图形界面开发框架）A cross-platform C++ GUI framework commonly used for robot operator-side software interfaces.
7.9Deep learning and robot-learning libraries
With data in hand, start learning: deep learning frameworks, model hubs, then LeRobot and reinforcement learning libraries.
- Today’s dominant deep-learning framework, used to train most embodied AI models.
- Google's NumPy-like framework for numerical computing and deep learning, with automatic differentiation and compilation.
- Google's open-source deep learning framework; robotics datasets often use its data format.
- MindSpore昇思 MindSporeHuawei's open-source deep-learning framework, built mainly to pair with Ascend chips.
- NVIDIA WarpWarpNVIDIA's Python framework that compiles ordinary Python functions into GPU code for writing simulations.
- Taichi Programming Language / QuadrantsTaichi（太极）/ QuadrantsA high-performance parallel programming language embedded in Python; the engine underneath the Genesis simulator.
- Hugging Face TransformersTransformers 库Hugging Face's open-source Python library for loading pretrained models with a few lines of code.
- Hugging Face DiffusersDiffusers 库Hugging Face's open-source diffusion-model toolkit, with ready-made schedulers, networks, and generation pipelines.
- safetensorssafetensors 权重格式A safe, fast-loading file format from Hugging Face for storing model weights.
- A model whose trained parameters are published for download, even if its training data and code aren't.
- The legal terms attached to public code or models that spell out who can use, modify, or sell them, and how.
- An unofficial mirror site that makes it fast to download Hugging Face models and datasets from inside China.
- ModelScope魔搭社区Alibaba's domestic open-source model and dataset community, roughly a Chinese counterpart to Hugging Face.
- LLaMA-Factory / ms-swift (ModelScope SWIFT)LLaMA-Factory / ms-swift（大模型与 VLM 微调框架）Two popular open-source fine-tuning frameworks for LLMs and VLMs, usable by just editing a config file.
- An open-source robot learning toolkit from Hugging Face, covering the whole pipeline from data collection to training to deployment.
- LeRobot EnvHubEnvHubLeRobot's mechanism for sharing simulation environments, publishing environment code to the Hugging Face Hub for one-line loading.
- Physical Intelligence's open-source code and weights repository for its π-series VLA models.
- Dexbotic (Dexmal VLA toolbox)DexboticAn open-source VLA training-and-deployment toolbox from Dexmal that brings several mainstream VLA methods into one codebase.
- An open-source codebase for assembling and training VLA models like building blocks.
- Google DeepMind's development kit for evaluating and fine-tuning its Gemini Robotics models.
- The standard interface library for reinforcement learning environments, and the successor to OpenAI Gym.
- A PyTorch-based library of classic reinforcement learning algorithms, with a simple API and reliable implementations.
- An open-source deep reinforcement-learning library that implements each algorithm as a single, readable file.
- ETH Zurich's open-source reinforcement learning library built for GPU-parallel simulation.
- A reinforcement learning training library optimized for massively parallel GPU simulation, common in the Isaac stack.
- A modular Python reinforcement learning library that plugs directly into Isaac Lab.
- HumanoidVerseHumanoidVerse（多仿真器人形 sim2real 框架）An open-source, multi-simulator humanoid robot learning framework from CMU's LeCAR Lab.
- Holosoma (Amazon FAR humanoid RL framework)Holosoma（亚马逊人形 RL 训练部署框架）An open-source humanoid reinforcement-learning framework for training and deployment, from Amazon's FAR team.
- GR00T-WholeBodyControlGR00T 全身控制代码库NVIDIA's open-source codebase for training, evaluating, and deploying whole-body controllers for humanoid robots.
7.10Training engineering and compute
Getting training running, then scaling it up: renting GPUs, tracking experiments, saving memory, multi-GPU clusters, and large-scale RL.
- AutoDLAutoDL（GPU 算力租用平台）A Chinese pay-by-the-hour GPU cloud platform, popular with students for running experiments.
- A local, browser-based tool for viewing training curves, images, and other logged data.
- Weights & Biases (W&B)Weights & BiasesA cloud platform for logging and comparing deep-learning training runs online.
- SwanLabSwanLab（国产训练实验追踪工具）An open-source Chinese experiment-tracking and visualization tool for model training, similar to W&B.
- Hydra (Meta configuration framework)Hydra 配置框架Meta's open-source Python configuration framework for composing YAML configs and overriding parameters from the command line.
- Numerical Precision Formats (FP32 / FP16 / BF16 / FP8 / INT8 / INT4)数值精度格式（FP32 / FP16 / BF16 / FP8 / INT8 / INT4）The bit width and format used to store a model's numbers, which sets its memory use, speed, and accuracy.
- A multi-GPU training method where each GPU holds a full copy of the model and processes a slice of the data, then syncs gradients.
- A data-parallel training method that splits a model's parameters, gradients, and optimizer state across multiple GPUs.
- Microsoft's open-source library for accelerating large-model distributed training, best known for its ZeRO memory optimizer.
- Hugging Face AccelerateAccelerate 库Hugging Face's training-helper library that lets the same PyTorch code run across multiple GPUs with minimal changes.
- Slurm Workload ManagerSlurm 集群调度The scheduling system that queues jobs and allocates GPUs on a shared compute cluster.
- RayRay 分布式框架Open-source distributed computing framework for scaling Python programs across many machines and GPUs.
- ByteDance's open-source large-model reinforcement learning framework, also used for VLA RL fine-tuning.
- Open-source, large-scale reinforcement learning training framework for embodied and agentic foundation models.
7.11Deployment and inference speedups
Moving a trained policy onto the robot: running it remotely or on-device, sped up with kernel optimization, export, and inference engines.
- Getting a trained model running on real hardware or a server so it outputs actions in real time.
- Running a large policy on a GPU server and having the robot send it observations and get actions back.
- Running model inference directly on a robot's own onboard computer, instead of calling a remote server.
- Splitting a robot's compute across the cloud, an edge server, and the robot itself, with each layer handling a different job.
- The basic computational building block of a neural network — like matmul or softmax — and its hardware implementation.
- NVIDIA's GPU-accelerated library of the operators most commonly used in deep learning.
- A GPU implementation of attention that reorders computation around memory hierarchy, giving identical results faster and with less memory.
- The one-line model-compilation acceleration API that PyTorch has offered since version 2.0.
- CUDA GraphsCUDA GraphRecording a sequence of GPU operations as a single graph, so it can be replayed in one submission with less launch overhead.
- TorchScript (torch.jit) / LibTorchTorchScript / LibTorch（策略导出与 C++ 部署）Exports a PyTorch model to a file that runs without Python, then loads it from C++.
- An open, general-purpose file format for neural networks that makes it easy to move models between frameworks and inference engines.
- Microsoft's open-source, cross-platform engine for running models saved in the ONNX format.
- NVIDIA TensorRTTensorRTNVIDIA's inference-acceleration library that compiles a trained model into a faster GPU runtime engine.
- NVIDIA JetPack SDKJetPackNVIDIA's official operating system and development kit for its Jetson embedded boards.
- NVIDIA TensorRT Edge-LLMTensorRT Edge-LLMNVIDIA's on-device inference framework for running large language and vision-language models on automotive and robotics chips.
- NVIDIA TensorRT-LLMTensorRT-LLMNVIDIA's open-source library that speeds up large language model inference on NVIDIA GPUs.
- An open-source, high-throughput inference and serving engine for large language models.
- Intel OpenVINOOpenVINOIntel's open-source inference-optimization and deployment toolkit for speeding up models on Intel CPUs, integrated graphics, and NPUs.
- Rockchip RKNN-ToolkitRKNN-ToolkitRockchip's model-conversion toolchain that turns neural networks into a format its NPUs can run.
- Horizon Robotics' toolchain for converting, quantizing, and deploying AI models on its own BPU chips.
- The low-level compute architecture and operator compilation toolchain for Huawei's Ascend NPUs.
7.12Industry platforms and ecosystems
Finally, how vendors package all these layers together: NVIDIA Isaac, Chinese robot operating systems, and cloud platforms.
- NVIDIA Isaac PlatformNVIDIA IsaacNVIDIA's umbrella brand for its full stack of robotics software, models, and tools.
- NVIDIA OmniverseOmniverseNVIDIA's 3D simulation and collaborative development platform, built around OpenUSD.
- NVIDIA Isaac ROSIsaac ROSA collection of GPU-accelerated ROS 2 packages from NVIDIA.
- NVIDIA's cloud orchestration platform that schedules robotics and physical-AI training, simulation, and data generation across different compute.
- An open-source project and consortium that brings ROS to industrial robots and manufacturing.
- TogetheROS.Bot (D-Robotics)TogetheROS.BotD-Robotics' robot software stack, built on top of and compatible with ROS 2.
- M-Robots OS (OpenHarmony-based Robot OS)M-Robots OSA Chinese robot operating system built on open-source HarmonyOS, focused on multi-robot collaboration.
- Link-U OS (AgiBot Lingqu OS)灵渠 OS（智元具身操作系统）AgiBot's operating system for embodied robots, with plans to open-source it in stages.
- AgiBot Genie StudioGenie Studio（智元具身开发平台）AgiBot's one-stop embodied-AI development platform, covering data collection, training, evaluation, and deployment.
- OM1OM1（OpenMind 机器人 OS）OpenMind's open-source, modular AI runtime that connects large language models to different robot bodies.
- BAAI's open-source cross-embodiment, multi-robot coordination framework built on a cloud brain, local cerebellum design.
- A general-purpose embodied-AI software platform from the Beijing Humanoid Robot Innovation Center.
- Huawei Cloud CloudRobo Embodied AI Platform华为云 CloudRoboHuawei Cloud's embodied-AI cloud platform, providing robots with large models and services hosted in the cloud.
- Tencent's open platform for embodied AI, offering foundation models and developer tools for robots.