Robots Atlas>ROBOTS ATLAS
UnifoLM (Unitree Foundational Model)

Control · Control & Planning

UnifoLM (Unitree Foundational Model)

0.8 (Research Preview)·Unitree Robotics

Beta API available
CATEGORYControl · Control & Planning
READINESSTRL 6
ADOPTION SCALEResearch / Prototype
LICENSESLicenseRef-Proprietary
FIRST RELEASE2024
UnitreeUnifoLMEmbodied AIRobot Foundation ModelWorld ModelVLAWMAHumanoid Robotics

UnifoLM (Unitree Foundational Model) is an internal foundation model developed by Unitree Robotics, officially announced at ChinaJoy 2024 and published in Unitree's Research section in 2025. UnifoLM is a Vision-Language-Action (VLA) model — it takes RGB images from the main camera, depth from a Livox LiDAR or RealSense, a language instruction (e.g., 'walk to the chair and sit down'), and emits a stream of low-level actions (joint torques, joint positions, foot placement).

The UnifoLM architecture is a transformer with three heads: a visual encoder (DINOv2 variant), a language encoder (Qwen-2 variant), and an action decoder (Diffusion Policy). Training ran on an NVIDIA HGX H100 cluster with synthetic Isaac Lab data and data collected from 100+ Unitree H1/G1 units during internal testing. The model is released as a pretrained checkpoint for external researchers.

UnifoLM is deployed onboard on NVIDIA Jetson AGX Orin with 20-40 ms action prediction latency. Used in Unitree's CES 2025 demonstrations (unboxing, card games, dancing). It competes with Figure Helix, Tesla Optimus Neural Network, and NVIDIA GR00T N1.

Type & Roles
Software types
Control Stack

A Control Stack is the set of software components responsible for control logic, motion planning, command execution, and coordination of a robot's actuators. It receives high-level goals or waypoints (from planning or AI layers) and translates them into low-level actuator commands, closed-loop feedback control, and safety checks. Common implementations include MoveIt (arm manipulation), Nav2 (mobile navigation), and custom whole-body control stacks for humanoids.

Perception Stack

A Perception Stack encompasses the software layers that process data from cameras, LiDARs, IMUs, microphones, and other sensors in order to recognise the surrounding environment, perform localisation, detect and track objects, and interpret the scene. It is typically the first processing stage in an autonomous robot's data pipeline, feeding its outputs to planning and control stacks.

Runtime

A Runtime is the environment or execution layer used to run code, load libraries, manage dependencies, and operate applications or services — either in real time or during normal system operation. In robotics this includes real-time operating system (RTOS) runtimes, ROS 2 executor runtimes, containerised execution environments (Docker, podman), and embedded C++ runtimes on microcontrollers.

Select an item to see its description.
Main category
Control & PlanningPerception & Vision Software
Roles in robotics ecosystem
PerceptionMotion Planning
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
Unitree SDK2
Maturity & Adoption
6 / 9
Demonstration phase
ResearchPrototypeProduction
Adoption scaleResearch / Prototype
Maintenance statusInternal / Proprietary – Not Public
First release2024
Last update15 April 2025
Deployments

Unitree H1 and G1 during CES 2025 and ChinaJoy 2024 demonstrations. Internal research at Unitree partner labs (Tsinghua, ShanghaiTech).

Community

No public community — the model is distributed under NDA to academic partners. A public release is announced for Q3 2025.

ROS supportCompatibility with ROS / ROS 2 ecosystem
Official Vendor ROS 2 WrapperOficjalny wrapper ROS 2 tworzony i utrzymywany przez producenta sprzętu lub oprogramowania
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 22.04JetPack Linux
Minimum hardware requirements
Minimum hardware requirements
CPUARM64 12-core (Jetson AGX Orin 64GB or Thor)
RAM (GB)32
GPUJetson Orin 275 TOPS (INT8) or Jetson Thor 2070 TFLOPS (FP8)
Disk (GB)50

Model ~6 GB FP16 / ~3 GB INT8. Inference 20-40 ms on Jetson AGX Orin 64 GB.

Packaging & distribution
Package managers
GitHub Releases / GitHub Actions ArtifactsDocker / Docker HubOTA (Over-The-Air Update)
CPU architectures
ARM64 / AArch64NVIDIA Jetson – AArch64 (JetPack)NVIDIA GPU (CUDA – x86_64)
Installation difficulty
LevelExpert only
Protocols and interfaces
Communication protocols
CycloneDDSDDS (Data Distribution Service)ROS 2 Topics
Hardware interfaces
Ethernet 1000BASE-T (Gigabit Ethernet)PCIe 4.0
Latency classes
Soft Real-Time (20–100 ms)
Deployment types
On RobotEdge
Supported simulators
NVIDIA Isaac Sim
NVIDIA Isaac Lab
MuJoCo
Official Docker images
unitreerobotics/unitree_ros2
Licenses
LicenseRef-ProprietaryProprietary – All Rights Reserved

License family: Proprietary – Commercial

Version history
0.8 (Research Preview)Apr 2025

Checkpoint released to academic partners.

0.5 (Beta)Jan 2025

CES 2025 demonstration: H1 and G1 performing manipulation tasks.

0.1 (Alpha)Aug 2024

First public announcement at ChinaJoy 2024.