Unitree G1
Bipedal humanoid robot by Unitree Robotics, designed as a compact research, development, and developer platform.
Applications
- Research
- Home Assistance
Active Open source Real-time capable API available
CATEGORYControl · Control & Planning
READINESSTRL 7
ADOPTION SCALEGrowing Community
LICENSESApache-2.0
FIRST RELEASE2025
NVIDIA GR00T N1 (Generalist Robot 00 Training N1) is a foundational AI model for humanoid robots, announced by NVIDIA at GTC 2025 in March 2025. It is the first open foundational model platform specifically designed for humanoid robots.
The GR00T N1 architecture is based on a dual-system approach inspired by Daniel Kahneman's System 1/2 theory. System 1 (fast): a neural motor policy (flow matching policy) generating action tokens at low latency — responsible for fluency and smooth movement. System 2 (slow): a large VLM (Llama-based or proprietary) responsible for understanding natural language, task sequence planning and contextual adaptation. Both systems operate asynchronously — System 2 sets high-level goals, System 1 executes them in real time.
GR00T N1 is trained on cross-embodiment data: a combination of data from many robot types (manipulators, mobile robots, humanoids), synthetic demonstrations from NVIDIA Isaac Sim, and wild-collected data. The data covers multiple modalities: RGB, depth, tactile (proprioception), language instructions. NVIDIA makes pre-trained weights available via Hugging Face and NGC (NVIDIA GPU Cloud).
GR00T N1 infrastructure is tightly integrated with the NVIDIA Isaac ecosystem: Isaac Sim for synthetic training data, Isaac Lab for RL fine-tuning, Isaac ROS for ROS 2 deployment. NVIDIA provides GR00T N1 as a base model for fine-tuning on specific hardware platforms (Unitree G1, Agility Digit, Figure 01/02, Boston Dynamics Atlas) through GR00T Isaac Lab workflows.
The model is open-weights (Apache 2.0 for weights and inference code), and the architecture and data are described in a technical report. This is a strategic NVIDIA initiative to position the Isaac platform as the standard for humanoid robotics AI, analogous to how CUDA became the standard for HPC.
Type & Roles
Software types
VLA / Foundation ModelRobot manufacturer SDK
API Library
An API Library is a software package that exposes programmatic interfaces for communicating with a device, service, or system. In robotics it typically forms a lightweight integration layer built on top of the manufacturer's official API or an open-source project, abstracting low-level protocol details and providing language-native bindings (Python, C++, Java, etc.).
Select an item to see its description.
Main category
Control & PlanningPerception & Vision SoftwareSDKsSimulation & Digital Twins
Roles in robotics ecosystem
Motion PlanningPerceptionDeveloper Enablement
Robot Learning
The robot learning role describes software for training a robot's control policies and manipulation/locomotion skills using machine learning methods. It covers: reinforcement learning in simulation with sim-to-real transfer, imitation learning and learning from demonstration, training Vision-Language-Action (VLA) models, and fine-tuning robotics foundation models. It typically uses massively parallelized simulation environments (Isaac Lab, MuJoCo) to generate training data, then deploys the trained policies on a physical robot.
Robot Control
Robot Control denotes the role of software responsible for motion control, command execution, coordination of actuating elements and the direct operational logic of the robot.
Select an item to see its description.
Software family
Family
NVIDIA Isaac
Maturity & Adoption
7 / 9
Prototype / pilot phase
ResearchPrototypeProduction
Adoption scaleGrowing Community
Maintenance statusActively Maintained
First release2025
Last update20 May 2026
Deployments
NVIDIA GTC 2025 demo — Jensen Huang demonstrated GR00T N1 on stage, controlling multiple humanoid models (Unitree G1, Figure 01/02, Agility Digit) performing manipulation tasks. Unitree G1 + GR00T — NVIDIA and Unitree Robotics announced a partnership; GR00T N1 fine-tuned on G1 hardware through the official Isaac Lab workflow. Unitree released Isaac Lab recipes for the G1. Figure AI — Figure 02 with a stack augmented by GR00T N1 pre-training (fine-tuned on Figure data). Demonstrations in BMW Spartanburg manufacturing work. Agility Robotics Digit — GR00T N1 as the base pre-trained model for Digit, fine-tuned by Agility for specific warehousing tasks. 1X Technologies NEO — Norwegian humanoid company integrating GR00T N1 as the model base for NEO Gamma. Labeling partners: Scale AI, Datagen — collecting and annotating cross-embodiment data for further GR00T training.
Community
GitHub: NVIDIA-ISAAC-ROS/isaac_ros_nitros + huggingface/groot-n1 — ~3 000 combined stars across related repositories. Active community via NGC Forums and Isaac ROS GitHub Discussions. NVIDIA Developer Forums: GR00T N1 thread with >500 posts in the first 3 months after announcement. Active response from the NVIDIA team. Citations: technical report 'GR00T N1: An Open Foundation Model for Generalist Humanoid Robots' (2025) — ~150 citations in 3 months. Partner ecosystem: 30+ robotics companies announced support for GR00T N1 (Unitree, Figure, Agility, Apptronik, Fourier, HEBI, Kepler, Skild AI and others). This is the largest partner coalition for a single robot foundation model. NGC Model Card: pre-trained model available for download via NVIDIA NGC with over 10 000 downloads in the first month.
Organizations
Integrates with
NI
NVIDIA Isaac ROS
NVIDIA Isaac ROS — collection of GPU-accelerated GEM packages for ROS 2: stereo depth, AprilTags, pose estimation, cuVSLAM, object detection on Jetson and x86 CUDA. Foundation of the NVIDIA robot perception stack.
NI
NVIDIA Isaac Lab
NVIDIA open framework for GPU-based robot learning, built on Isaac Sim. Trains RL policies across thousands of parallel environments. Successor to Isaac Gym and Orbit; Apache 2.0 since 2024.
NI
NVIDIA Isaac Sim
Photorealistic robotics simulator (RTX) with advanced PhysX 5 physics. Built on Omniverse Kit, supports ROS 2, synthetic data generation (SDG), Isaac Lab training, and the Isaac ROS deployment pipeline on Jetson.
R2
ROS 2
Open-source framework for building robot software. The successor to ROS 1, built on DDS with native support for distributed, real-time and multi-platform systems. The de facto standard in research and commercial robotics.
R2
ROS 2 Humble Hawksbill
LTS release of the ROS 2 framework based on Ubuntu 22.04, supported through May 2027. The most widely deployed release in humanoids, AMRs, and research platforms. Full integration with Nav2, MoveIt 2 and ros2_control.
R2
ROS 2 Jazzy Jalisco
ROS 2 Jazzy Jalisco — the latest LTS (May 2024, supported until May 2029) on Ubuntu 24.04. C++17, Fast DDS 3.0, improved executors, RMW Zenoh, better QoS for variable networks.
M2
MoveIt 2
Open-source motion planning, manipulation, and kinematics framework for ROS 2 (Foxy → Jazzy). Stewarded by PickNik Robotics. The de facto standard for manipulators in the ROS ecosystem.
L
LeRobot
LeRobot (Hugging Face) — open-source robot learning framework with ACT, Diffusion Policy and TDMPC implementations. Standardises data collection and policy training for manipulators and mobile robots.
Related robotics software
Π(
π0 (pi-zero)
Physical Intelligence's first 'generalist robot policy' — VLA with flow matching, 50 Hz actions, trained on 10,000+ hours of demonstrations. Open weights π0-base (February 2025, Apache 2.0).
Π(
π0.5 (pi-zero-5)
π0.5 (pi-zero-5, Physical Intelligence) — evolution of π0 focused on open-world mobile manipulation: generalises to new environments and tasks without additional fine-tuning thanks to large-scale training data.
R(
RT-2 (Robotics Transformer 2)
A Google DeepMind Vision-Language-Action model based on PaLI-X / PaLM-E. Translates images + language into robot action tokens. The first 'language to action' at real-robot scale (2023).
O
OpenVLA
Open-source replication of RT-2 (Stanford + Berkeley + TRI, June 2024). A 7B-parameter VLA (Llama 2 + DINOv2 + SigLIP), trained on 970k demonstrations from Open X-Embodiment.
L
LeRobot
LeRobot (Hugging Face) — open-source robot learning framework with ACT, Diffusion Policy and TDMPC implementations. Standardises data collection and policy training for manipulators and mobile robots.
NI
NVIDIA Isaac ROS
NVIDIA Isaac ROS — collection of GPU-accelerated GEM packages for ROS 2: stereo depth, AprilTags, pose estimation, cuVSLAM, object detection on Jetson and x86 CUDA. Foundation of the NVIDIA robot perception stack.
NI
NVIDIA Isaac Lab
NVIDIA open framework for GPU-based robot learning, built on Isaac Sim. Trains RL policies across thousands of parallel environments. Successor to Isaac Gym and Orbit; Apache 2.0 since 2024.
NI
NVIDIA Isaac Sim
Photorealistic robotics simulator (RTX) with advanced PhysX 5 physics. Built on Omniverse Kit, supports ROS 2, synthetic data generation (SDG), Isaac Lab training, and the Isaac ROS deployment pipeline on Jetson.
Supported robot models
Unitree H1
Full-size general-purpose humanoid (~180 cm, ~47 kg). Bipedal, 5 DOF per leg, 3.3 m/s walking speed, 360° perception via 3D LiDAR + depth camera, Unitree M107 PMSM joint motors (~360 N·m peak knee torque).
Applications
- Research
Figure 03
Figure 03 is the third-generation humanoid robot from Figure AI, designed for Helix, home environments, and scalable mass production.
Applications
- Factory Automation
- Industrial Logistics
- Production Line Operation
- Warehouse Automation
- Object Manipulation
- Home Assistance
Tesla Optimus
General-purpose humanoid robot developed by Tesla since 2021 (Tesla Bot). About 173 cm tall, 57 kg, 20 kg payload. Generation 3 hands have 22 degrees of freedom. Controlled by an AI system derived from Tesla FSD.
Applications
- Factory Automation
- Production Line Operation
- Object Manipulation
- Home Assistance
Atlas
Boston Dynamics bipedal humanoid robot. The fully electric generation unveiled in 2024 succeeds the hydraulic Atlas that was retired after more than a decade of research.
Applications
- Factory Automation
- Research
- Production Line Operation
- Object Manipulation
MagicBot Z1
Compact, high-dynamics bipedal humanoid by MagicLab. 140 cm, 40 kg, 24–50 DOF, walking speed up to 2.5 m/s. Unveiled on 8 July 2025 with martial arts and acrobatic demos.
Applications
- Research
- Hospitality
- Retail Trade
- Home Assistance
- Object Manipulation
Target robotic platforms
Humanoid
Robotic Arm
Mobile Robot
Research Robot
Quadruped
Supported hardware
NVIDIA Jetson AGX Orin 64GB
NVIDIA Jetson AGX Thor
compute · compute_modules · industrial
NVIDIA Jetson Orin NX 16GB
compute · compute_modules · industrial
Intel RealSense D435i
sensing · cameras · research · Stereo RGB-D camera
Intel RealSense D455
sensing · cameras · professional
Stereolabs ZED 2i
sensing · cameras · industrial
Stereolabs ZED X
sensing · cameras · industrial
ROS supportCompatibility with ROS / ROS 2 ecosystem
Official Vendor ROS 2 WrapperOficjalny wrapper ROS 2 tworzony i utrzymywany przez producenta sprzętu lub oprogramowania
Official ROS 2 PackagePakiet dostępny w oficjalnym rejestrze ROS 2 przez rosdep / apt (packages.ros.org)
ROS 2 Component (Composable Node)Węzeł ROS 2 zarejestrowany jako composable node (component) dla intraprocess communication
System capabilities
⊙
Open source
Source code is publicly available under an open-source license — enables security audits, custom modifications, and integration without licensing barriers.
✓
⚡
Real-time capable
Designed with timing-determinism guarantees — meets the requirements of control loops, safety systems, and tasks demanding low, predictable latency.
✓
⟨/⟩
API available
The software exposes a programmable interface (REST, gRPC, SDK, or language bindings) that enables automation and integration with other systems.
✓
📦
Pre-built / binary
Distributed as ready-to-use binary packages, container images, or installers — no need to build from source.
✓
Programming languages
PythonC++CUDA
Operating systems
Ubuntu 22.04Ubuntu 20.04JetPack Linux
Ubuntu 24.04
Ubuntu 24.04 LTS 'Noble Numbat' — supported until April 2029. The host for ROS 2 Jazzy.
Select an item to see its description.
Minimum hardware requirements
Minimum hardware requirements
CPUOn-robot inference: Jetson AGX Thor (128 TOPS) or Jetson AGX Orin 64GB for lighter variants. Training: GPU cluster (DGX A100/H100). Minimum 16-core x86 CPU for local fine-tuning.
RAM (GB)Inference: 32 GB RAM minimum (64 GB zalecane). Fine-tuning lokalnie: 128 GB+ RAM z A100 80 GB VRAM.
GPUInference: NVIDIA RTX 4090 (24 GB) or Jetson AGX Thor minimum. Fine-tuning: A100 40/80 GB or H100. GR00T N1 optimised for TensorRT for Jetson deployment.
Disk (GB)Model weights ~10-40 GB (zależnie od wariantu). Isaac Sim: min 50 GB. Pełny dataset syntetyczny może wymagać >5 TB NVMe.
Recommended deployment in NGC Docker container (nvidia-isaac-ros:latest). Requires JetPack 6.0+ on Jetson. Proprioception: IMU + joint encoders required for full modality.
Packaging & distribution
Package managers
pip / PyPIDocker – NVIDIA NGC RegistryGitHub Releases / GitHub Actions ArtifactsSource – Python (setup.py / pyproject.toml)conda / mamba
CPU architectures
x86_64 (AMD64)NVIDIA GPU (CUDA – x86_64)NVIDIA Jetson – AArch64 (JetPack)ARM64 / AArch64
Installation difficulty
LevelAdvanced
Protocols and interfaces
Communication protocols
ROS 2 TopicsDDS (Data Distribution Service)gRPCShared Memory (POSIX / mmap)
Hardware interfaces
Ethernet 1000BASE-T (Gigabit Ethernet)Ethernet 10GBASE-T (10 Gigabit Ethernet)USB 3.0 / 3.1 Gen 1MIPI CSI-2PCIe 4.0
Latency classes
Soft Real-Time (20–100 ms)Soft Real-Time (100–500 ms)Hard Real-Time (5–20 ms)
Deployment types
On RobotCloudEdgeContainerizedLocal Workstation
Supported simulators
NVIDIA Isaac Sim
NVIDIA Isaac Lab
MuJoCo
Gazebo Harmonic
Official Docker images
nvcr.io/nvidia/isaac/rosnvcr.io/nvidia/isaac-sim
Licenses
Apache-2.0Apache License 2.0v2.0
License family: Permissive
ModificationDistributionCommercial useSublicensingPrivate useROS-compatibleOSI approvedFSF Free/LibreRequires attributionPatent grant
Version history
N1.5Sept 2025
GR00T N1.5 — improved model with expanded cross-embodiment dataset, better natural-language instruction understanding and TensorRT 10 optimisation for Jetson.
N1Mar 2025
First public release of GR00T N1 at GTC 2025. Open-weights (Apache 2.0), Isaac Sim/Lab/ROS integration, dual-system architecture.