Robot API and SDK Directory (2026)
The robot development tooling landscape is confusing. There are platform SDKs (NVIDIA Isaac), robot operating systems (ROS 2), manufacturer SDKs (Unitree, UFACTORY), AI training frameworks (LeRobot, Isaac Lab), and cloud platforms. This directory lists every major option, what it does, and when you need it.
Use this as a reference when starting a new robotics project. Each entry tells you what the tool does, what license it uses, and where to get it.
For the model and data layers that sit above these SDKs, use the robot foundation models guide, open model comparison, and robotics dataset and benchmark directory.
Platform-level SDKs
These provide comprehensive robot development environments, not just a single library.
NVIDIA Isaac
| Aspect | Details |
|---|---|
| What | Full-stack robot development platform: simulation, training, deployment |
| Components | Isaac Sim, Isaac Lab, Isaac ROS, Isaac GR00T, Cosmos |
| License | Free for development (GPU hardware required) |
| Languages | Python, C++ |
| Hardware req. | NVIDIA GPU (RTX 3070+ for Sim, Jetson for deployment) |
| Link | developer.nvidia.com/isaac |
| Best for | Teams building serious AI-powered robots with NVIDIA hardware |
The most comprehensive platform available. Covers everything from simulation through training to edge deployment. Overkill for hobby projects, essential for production robots. See our NVIDIA Isaac vs Cosmos guide for how the components fit together.
Isaac ROS 5.0 status: NVIDIA’s release notes date version 5.0.0 to September 21, 2026; NVIDIA announced it publicly on September 22. The available open-source release moves the supported distribution to ROS 2 Lyrical on Ubuntu 24.04 and rebuilds accelerated message transport around standard ROS messages using rosidl::Buffer with a CUDA backend.
The advertised “AI agent” capability is a set of instructions in the open Agent Skills format for compatible coding assistants, not an autonomous robot agent. Isaac ROS CLI ships isaac-ros-activate to activate the development environment and an early-access migrate-node-to-rosidl-buffer skill to assist source migration from direct NITROS APIs. NVIDIA’s broader skills catalog includes additional Physical AI workflows, but installing a skill does not independently deploy code to a robot, validate generated code, orchestrate Isaac Sim, or establish runtime autonomy. Developers retain responsibility for tool permissions, review, build, simulation and hardware testing.
Migration is not entirely drop-in: 5.0 removes several direct NITROS packages and types, retains isaac_ros_nitros_bridge_ros2 as deprecated, and requires source changes where applications called those APIs directly. Platform compatibility still needs to be checked per package and Jetson configuration.
AWS Physical AI Toolchain
Reference implementation checked October 12, 2026: the AWS-maintained sample repository integrates AWS infrastructure with NVIDIA tooling. Public sample code and deployment modules are available; that does not make the whole robot lifecycle production-ready.
| Component | Repository status | Deployment scope |
|---|---|---|
| Foundation infrastructure | Available | IaC for shared storage, containers, IAM and networking |
| Cosmos | Available | World generation and Transfer 2.5 augmentation modules |
| Isaac Lab | Available | RL training on SageMaker and AWS Batch |
| Isaac GR00T | Available | N1.6 fine-tuning recipe; not a claim about every GR00T release |
| Isaac Sim | Available | GPU simulation workstation module |
| OSMO | Available | 6.3 orchestration on EKS |
| Isaac Lab Arena evaluation | Preview | Closed-loop evaluation and checkpoint-registration gates |
| Jetson edge deployment | Planned | Packaging through EKS Hybrid Nodes and Greengrass; not an available end-to-end module |
The sample code is Apache 2.0; NVIDIA components, models and datasets retain their own terms. Available means the repository supplies the named reference module, not independent hardware validation or a managed-service SLA. Its high-level deployment diagram includes edge delivery, but the component table still labels that module Planned. Start with the component README, GPU quotas and teardown instructions, then validate your own sim-to-real workflow. The toolchain reduces infrastructure assembly work; it does not supply finished robot behavior.
Hugging Face LeRobot
| Aspect | Details |
|---|---|
| What | Open-source framework for robot learning: data collection, training, deployment |
| Components | Teleoperation, dataset management, imitation learning, pre-trained policies |
| License | Apache 2.0 |
| Languages | Python |
| Hardware req. | Any computer (CPU works, GPU recommended for training) |
| Link | github.com/huggingface/lerobot |
| Best for | Researchers, learners, anyone starting with affordable robot arms |
The fastest way to go from zero to autonomous robot manipulation. Designed around the SO-101 arm but supports other hardware. Includes pre-trained VLA models (ACT, pi-zero, SmolVLA) that work out of the box. See our Best Affordable Robot Arms guide for compatible hardware.
ROS 2 (Robot Operating System)
| Aspect | Details |
|---|---|
| What | Middleware framework for robot software: communication, coordination, tools |
| Components | pub/sub messaging, tf2 (transforms), nav2 (navigation), MoveIt2 (manipulation) |
| License | Apache 2.0 |
| Languages | Python, C++ |
| Hardware req. | Linux (Ubuntu recommended), some Windows/macOS support |
| Link | ros.org |
| Best for | Any multi-component robot system that needs standardized communication |
The industry standard for robot software integration. Not an AI framework, a communication and coordination layer that everything else plugs into. Most manufacturer SDKs offer ROS 2 bridges. Most NVIDIA Isaac tools integrate with ROS 2 natively.
MoveIt 2
| Aspect | Details |
|---|---|
| What | ROS 2 manipulation and motion-planning framework |
| Components | Planning pipelines, collision checking, kinematics, MoveIt Servo, ros2_control integration |
| License | BSD 3-Clause |
| Stewardship | Open-source community with PickNik as a major steward |
| Ownership event | Qualcomm agreement to acquire PickNik announced September 23, 2026; pending close |
| Planned integration | Optimized Dragonwing and Arduino paths; not yet general availability |
| Link | moveit.ai |
Qualcomm and PickNik say MoveIt 1 and 2 will remain open source, hardware agnostic and community driven, with third-party hardware support continuing. The acquisition has not closed, and the announced Dragonwing/Arduino integration should not be described as an already-shipping dependency.
Manufacturer SDKs
These come from robot hardware companies and provide control over their specific platforms.
Unitree SDK 2
| Aspect | Details |
|---|---|
| What | Official SDK for Unitree robots (G1, H2, Go2, B2) |
| License | Proprietary (free to use with Unitree hardware) |
| Languages | Python, C++ |
| Link | github.com/unitreerobotics/unitree_sdk2 |
| Features | Low-level motor control, high-level locomotion API, ROS 2 bridge |
| Best for | Anyone building on Unitree hardware |
Also notable: Unitree maintains a LeRobot integration (unitree_IL_lerobot) for using Hugging Face’s imitation learning framework directly with G1’s dual-arm dexterous hands.
UFACTORY xArm SDK
| Aspect | Details |
|---|---|
| What | Official SDK for UFACTORY xArm robot arms |
| License | BSD (open-source) |
| Languages | Python, C++, ROS 2 |
| Link | github.com/xArm-Developer |
| Features | Joint/Cartesian control, force sensing, gripper control, collision detection |
| Best for | Anyone using xArm hardware for research or production |
Dobot SDK
| Aspect | Details |
|---|---|
| What | Official SDK for Dobot robot arms (Nova, CR, MG400) |
| License | Proprietary (free with Dobot hardware) |
| Languages | Python, C++, C#, Java |
| Link | github.com/Dobot-Arm |
| Features | TCP/IP control, graphical programming (Dobot+), ROS integration |
| Best for | Education, light automation with Dobot hardware |
Robotiq C++ SDK and ROS 2 stack
| Aspect | Details |
|---|---|
| What | Vendor-maintained gripper SDK plus ROS 2 ros2_control integration |
| Models | 2F-85, 2F-140 and Hand-E through different support paths |
| Transport | Modbus RTU over the SDK’s serial link |
| Link | github.com/robotiq/ros |
| Best for | Integrating Robotiq adaptive grippers into ROS 2 manipulation stacks |
The official repository ships descriptions for the 2F-85 and 2F-140. Robotiq says hardware validation to date is on the 2F-85; the bundled 2F-140 description has not yet been hardware validated. The driver is model-agnostic enough to work with Hand-E when a suitable URDF is supplied, but a Hand-E description is not bundled. Do not generalize 2F-85 validation across the family.
The same stack exposes standard joint interfaces for topic-based simulation, including an Isaac Sim-compatible path. Simulation compatibility does not prove real-hardware behavior. The official code continues PickNik’s earlier ROS 2 package while replacing its in-tree Modbus layer with Robotiq’s C++ SDK.
AI and Training Frameworks
These provide the machine learning infrastructure for training robot policies.
NVIDIA Isaac Lab
| Aspect | Details |
|---|---|
| What | GPU-accelerated framework for robot reinforcement and imitation learning |
| License | MIT (open-source) |
| Languages | Python |
| Hardware req. | NVIDIA GPU (runs thousands of parallel environments) |
| Link | github.com/isaac-sim/IsaacLab |
| Best for | Training locomotion and manipulation policies at scale |
NVIDIA Isaac GR00T
| Aspect | Details |
|---|---|
| What | End-to-end humanoid development platform with pre-trained VLA model |
| License | Open (specific terms per version) |
| Current version | GR00T 1.7 (Cosmos-Reason2-2B backbone, 32K hours real + 8K hours sim training) |
| Link | developer.nvidia.com/isaac/groot |
| Best for | Humanoid robot developers who want a ready-made VLA foundation |
MuJoCo
| Aspect | Details |
|---|---|
| What | Physics simulator optimized for robot control and reinforcement learning |
| License | Apache 2.0 (open-source since 2022, originally DeepMind) |
| Languages | Python (via mujoco-py), C |
| Link | mujoco.org |
| Best for | RL researchers, fast physics simulation, contact-rich manipulation |
Lighter and faster than Isaac Sim but less photorealistic. Often used for rapid policy iteration before transferring to Isaac Sim for visual fidelity testing.
PyBullet
| Aspect | Details |
|---|---|
| What | Open-source physics engine for robot simulation and RL |
| License | zlib (very permissive) |
| Languages | Python, C++ |
| Link | pybullet.org |
| Best for | Quick prototyping, academic research, when you need something simpler than Isaac Sim |
Older and less maintained than MuJoCo or Isaac Sim, but still widely used in academic RL research due to simplicity and good documentation.
Cloud and Deployment
Hesai Kosmo: capture and reconstruction
Evidence checked October 15, 2026: Hesai’s October 14 announcement describes a real-world capture → reconstruction → simulation-data stack, not a foundation model or robot controller.
| Field | Current evidence |
|---|---|
| Kosmo One | Announced/demonstrated spatial camera: lidar, one-inch imaging sensor and IMU |
| Kosmo World | Cloud reconstruction/data-generation platform; 3D Gaussian splats, meshes, depth and semantic segmentation |
| Resolution | Maximum 8K capture; maximum 6K user export, per vendor footnote |
| Ecosystem | Hesai names Isaac Sim, Isaac Lab, Unreal Engine, RenderMan and Houdini integrations; per-version shipping connectors/compatibility not independently verified |
| Demonstrations | Launch scanning and vendor-reported Sharpa/Dairy Queen workflow; RenderMan workflow/resources linked from announcement |
| Availability / region / price | Launched with vendor-described customer cases; public price, self-service developer release, regional ordering and general availability UNKNOWN in cited source |
| License / access | Vendor describes raw-data/asset access; no blanket open-source license or unrestricted dataset license established |
Do not infer a generally shipping developer SDK from an integration list. Request supported formats, connector versions, terms, lead time and regional access. Validate exported geometry and dynamics through the Real2Sim workflow before robot training; vendor time savings and success rates are not our measurements.
Arm AI Portal
| Aspect | Details |
|---|---|
| What | Model discovery, performance data, deployment code, workflows, and optimization resources for supported Arm targets |
| Scope | Cloud, mobile, edge, embedded, and physical AI; broader than robotics |
| Launch status | Available since September 8, 2026; bring-your-own-model optimization and some agent-ready resources have separate early-access boundaries |
| Initial model/runtime examples | Qwen, Gemma, and Ultralytics YOLO resources using ExecuTorch, LiteRT, and ONNX Runtime |
| Agent access | Portal resources can be queried through Arm’s documented MCP endpoint |
| Link | developer.arm.com/ai |
| Best for | Teams selecting and deploying supported vision or AI models on Arm-based edge hardware |
The portal is a developer-resource catalog, not a robotics framework and not proof that every listed model performs well on every Arm device. Check the model’s exact target, runtime, precision, memory use, and published performance data before choosing it for a robot. Arm says its portal resources are machine-discoverable for coding agents, while some optimization and agent-ready capabilities remain early access.
NVIDIA Jetson (edge deployment)
| Aspect | Details |
|---|---|
| What | Edge AI compute modules for deploying trained models on real robots |
| Products | Jetson Orin Nano ($249), Jetson Orin NX ($599), Jetson Thor (upcoming) |
| Best for | Running trained policies on-robot without cloud connectivity |
Jetson Thor is specifically designed for humanoid robots, providing the onboard compute for running GR00T policies locally.
Foxglove
| Aspect | Details |
|---|---|
| What | Visualization and debugging platform for robotics data |
| License | MIT (core), commercial (cloud features) |
| Link | foxglove.dev |
| Best for | Debugging robot behavior, visualizing sensor data, log analysis |
The modern replacement for RViz (ROS’s built-in visualizer). Web-based, supports ROS 2 natively, and handles large datasets better.
Decision matrix
| Your situation | Primary tool | Supporting tools |
|---|---|---|
| Learning physical AI from scratch | LeRobot + SO-101 | ROS 2 (later) |
| Building a humanoid with Unitree hardware | Isaac GR00T + Unitree SDK | Isaac Sim, ROS 2 |
| Training manipulation policies | Isaac Lab | MuJoCo (prototyping), Isaac Sim (validation) |
| Deploying perception on a mobile robot | Isaac ROS | ROS 2, Jetson |
| Building a product with an xArm | xArm SDK + ROS 2 | Isaac Sim (for testing) |
| Academic RL research | MuJoCo or PyBullet | Isaac Lab (for scale) |
FAQ
Do I need ROS 2 for every robot project?
No. ROS 2 is most valuable when your robot has multiple subsystems that need to communicate (cameras, arms, navigation, planning). For a single robot arm running a learned policy, LeRobot or a manufacturer SDK is simpler. ROS 2 becomes essential as system complexity grows.
Is NVIDIA Isaac free?
Yes for development. All Isaac components (Sim, Lab, ROS, GR00T models) are free to download. You pay only for NVIDIA GPU hardware. There are no per-robot licensing fees for deployment.
Can I mix tools from different ecosystems?
Yes. ROS 2 is the common integration layer. A typical project might use: Isaac Sim for simulation, LeRobot for data collection, Isaac Lab for training, Unitree SDK for hardware control, and ROS 2 to connect everything. Cross-ecosystem interoperability is the norm, not the exception.
What if I do not have an NVIDIA GPU?
LeRobot works on CPU (slower training but functional). MuJoCo and PyBullet work without NVIDIA hardware. Isaac Sim and Isaac Lab require NVIDIA GPUs. For serious physical AI development, an NVIDIA GPU is a near-requirement in 2026.
Which tool should I learn first?
If you are new to robotics: LeRobot (immediate hands-on results). If you already know robotics and want to add AI: Isaac Lab. If you are building a production system: ROS 2 first, then layer in AI tools.
Sources
- NVIDIA Isaac Developer Platform — full-stack robot development platform
- Hugging Face LeRobot GitHub — open-source robot learning library (Apache 2.0)
- ROS 2 official site — middleware framework for robot communication
- Unitree SDK 2 GitHub — official SDK for Unitree robots
- UFACTORY xArm Developer GitHub — BSD-licensed xArm SDKs
- NVIDIA Isaac GR00T GitHub — humanoid foundation model
- MuJoCo official documentation — physics simulator (Apache 2.0)
- PyBullet official site — lightweight physics simulation
- Foxglove developer tools — robot data visualization
- NVIDIA Isaac Lab GitHub — GPU-accelerated training framework
- Arm AI Portal — official model, performance, deployment, optimization, and MCP resource portal
- Arm AI Portal launch announcement — launch scope and early-access boundaries
- NVIDIA: Isaac ROS 5.0 — current availability, Lyrical/Ubuntu support, agent-ready workflows and Jetson range
- NVIDIA Isaac ROS 5.0 release notes — exact release date, ROS 2 migration, package changes and bundled Agent Skills
- NVIDIA Isaac ROS getting started — supported coding-assistant workflow and skill installation model
- Qualcomm: agreement to acquire PickNik — transaction and MoveIt commitments
- Robotiq official ROS repository — vendor-maintained SDK, model descriptions, hardware-validation boundary and simulation interfaces