Unitree G1
Last reviewed
Sources
30 citations
Review status
Source-backed
Revision
v7 ยท 5,813 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
30 citations
Review status
Source-backed
Revision
v7 ยท 5,813 words
Add missing citations, update stale details, or suggest a clearer explanation.
| Unitree G1 | |
|---|---|
 | |
| General information | |
| Manufacturer | Unitree Robotics |
| Country of origin | China |
| Year unveiled | 2024 |
| Status | In production |
| Price | From $16,000 USD (base) |
| Availability | Commercially available |
| Website | unitree.com/g1 |
The Unitree G1 is a compact, general-purpose humanoid robot developed by Unitree Robotics, a Chinese robotics company headquartered in Hangzhou, Zhejiang. Unveiled on May 13, 2024, at the IEEE International Conference on Robotics and Automation (ICRA) in Yokohama, Japan, the G1 sells from a base price of $16,000, making it one of the most affordable production humanoid robots in the world [16]. It stands about 127 cm (4 ft 2 in) tall, weighs roughly 35 kg (77 lb), walks at up to 2 m/s, and ships with 23 to 43 degrees of freedom depending on configuration. Significantly smaller and lighter than its full-size sibling the Unitree H1, the G1 is designed as an accessible platform for research, education, and light commercial applications.
The G1 is available in multiple configurations ranging from a base model with 23 degrees of freedom to advanced EDU variants with up to 43 degrees of freedom, optional dexterous hands, and an NVIDIA Jetson Orin AI co-processor. The robot can walk at speeds up to 2 m/s (4.5 mph), perform dynamic movements such as standing side flips, and fold down to 690 mm for easy transport. By the end of 2025, Unitree had shipped more than 5,500 humanoid robots, accounting for roughly 32.4 percent of the estimated 17,000 humanoid units delivered globally that year, the largest share of any maker and enough to place Unitree first worldwide in humanoid robot shipments [3]. The G1 accounted for a significant portion of that volume and became the de facto default research platform at universities with humanoid programs.
At launch, Unitree positioned the robot as a "Humanoid Agent AI Avatar" that "has flexibility beyond ordinary people and unlocks unlimited movement potential," able to "perform high-load dynamic movements, such as dynamic stand-up, seat folding, dance stick, etc." [4][26]
Unitree Robotics (formally Hangzhou Yushu Technology Co., Ltd.) was founded by Wang Xingxing in August 2016. Wang, who was born in 1990 in Ningbo, Zhejiang, developed a prototype quadruped robot called XDog during his graduate studies at Shanghai University in 2015. The design used low-cost brushless motors rather than expensive hydraulic actuators, attracting early investors and laying the foundation for Unitree's strategy of making advanced robots affordable.
The company released its first commercial quadruped robot, Laikago, in 2017. Subsequent products, including the Go1 (2021) and Go2, established Unitree as the global market leader in quadruped robotics, with an estimated 60 to 70 percent of worldwide quadruped shipments [1]. The company entered the humanoid robot market with the H1 in August 2023, followed by the G1 in May 2024 and the consumer-oriented R1 ($5,900) in 2025.
Unitree's investors include Tencent, Alibaba, China Mobile, Ant Group, Geely Capital, and HongShan Capital (formerly Sequoia China). The company raised its Series C round in June 2025 at a valuation of approximately $1.7 billion [2]. In March 2026, Unitree filed for an initial public offering on the Shanghai Stock Exchange's STAR Market, seeking to raise 4.2 billion yuan (roughly $610 million) at a reported target valuation of $7 billion [3].
The G1 was formally introduced on May 13, 2024, during ICRA 2024 in Yokohama, Japan. Unitree described it as a "Humanoid Agent AI Avatar" and positioned it as the company's second humanoid platform after the H1 [4]. The choice to launch at ICRA, the premier academic robotics conference, signaled that researchers and educators were the primary initial audience.
At launch, the G1's base price of $16,000 (excluding tax and shipping) was roughly one-sixth the cost of the H1 ($90,000 to $150,000) and a fraction of the price of competing platforms from Boston Dynamics, Agility Robotics, and Figure AI, which typically cost $150,000 to $300,000 or are not sold to external buyers at all [16][26]. The aggressive pricing drew widespread media coverage and positioned the G1 as a potential turning point in making humanoid robotics accessible beyond well-funded corporate and government laboratories.
Following the initial launch, Unitree transitioned the G1 into mass production, with output beginning on August 20, 2024, after hardware optimizations for large-scale manufacturing [5]. In 2025, the company introduced the G1 EDU line, a series of developer-focused configurations with enhanced computing, expanded degrees of freedom, and full software development kit support. By 2026, Unitree offered 16 distinct G1 configurations spanning prices from approximately $17,990 to $73,900, covering everything from basic demonstration units to fully equipped research platforms [22].
The G1 stands 1,320 mm tall, 450 mm wide, and 200 mm deep in its upright posture. It weighs approximately 35 kg with the battery installed. Each leg measures roughly 600 mm in length, and the arm span is approximately 450 mm.
A key design feature is the G1's foldable frame. The robot can collapse into a seated position measuring just 690 x 450 x 300 mm, small enough to fit inside a large suitcase or standard transport case. This portability is a practical advantage for researchers moving the robot between laboratories, conferences, and demonstration venues.
| Specification | Value |
|---|---|
| Height (standing) | 1,320 mm (127 cm / 4 ft 2 in) |
| Width | 450 mm |
| Depth | 200 mm |
| Height (folded) | 690 mm |
| Folded dimensions | 690 x 450 x 300 mm |
| Weight (with battery) | ~35 kg (77 lb) |
| Leg length | ~600 mm |
| Arm span | ~450 mm |
The G1 is available in multiple configurations with varying degrees of freedom. The base model provides 23 DOF, distributed as follows: 6 DOF per leg (3 hip, 1 knee, 2 ankle), 5 DOF per arm (3 shoulder, 1 elbow, 1 wrist), and 1 DOF at the waist. Higher-tier configurations add wrist rotation axes, additional waist joints, and dexterous hands, expanding the total to 43 DOF.
| Configuration | Total DOF | Waist DOF | Arm DOF (per arm) | Hand | Notes |
|---|---|---|---|---|---|
| G1 Standard (Base) | 23 | 1 | 5 | Gripper (no fingers) | Entry-level model |
| G1 EDU Standard (U1) | 23 | 1 | 5 | Gripper | Adds Jetson Orin, full SDK |
| G1 EDU Plus (U2) | 29 | 3 | 5 + 2 wrist axes | Gripper | Expanded waist and wrist |
| G1 EDU Ultimate A (U3) | 43 | 3 | 5 + 2 wrist + Dex3-1 hand (7 DOF) | 3-finger dexterous | Maximum DOF with 3-finger hands |
| G1 EDU Ultimate D (U5) | 41 | 3 | 5 + 2 wrist + Dex5-1 hand | 5-finger dexterous | Up to 94 tactile sensors per hand |
The Dex3-1 three-fingered hand provides 7 DOF per hand (3 DOF thumb, 2 DOF index finger, 2 DOF middle finger) with force-position hybrid control that enables precise grasping and manipulation of everyday objects such as bottles, tools, and door handles [6]. The five-fingered Dex5-1 variant, available on higher-tier Ultimate configurations, offers 20 DOF per hand (16 active joints and 4 passive joints), up to 4.5 kg of gripping force, and approximately 1 mm positioning accuracy. Each Dex5-1 hand can be specified with up to 94 miniature tactile sensors embedded across the fingertips, finger pads, and palm, supplying haptic feedback for advanced manipulation research and dexterity studies [15].
The G1 uses Unitree's proprietary permanent magnet synchronous motor (PMSM) joint actuators. These low-inertia motors feature hollow shafts for internal cable routing, dual encoders per joint for precise position and velocity feedback, and local air cooling for thermal management. Unitree states that the maximum joint torque can reach 120 N.m on higher-tier configurations [26].
| Parameter | G1 Standard | G1 EDU |
|---|---|---|
| Maximum knee torque | 90 N.m | 120 N.m |
| Arm payload capacity | ~2 kg | ~3 kg |
| Waist rotation range | +/-155 degrees | +/-155 degrees |
| Knee flexion range | 0 to 165 degrees | 0 to 165 degrees |
| Hip range | +/-154 to 170 degrees | +/-154 to 170 degrees |
The EDU variant's higher knee torque (120 N.m versus 90 N.m on the standard model) provides more powerful and stable locomotion, particularly during dynamic movements such as stair climbing, getting up from the ground, and carrying objects [27].
The G1 features a multi-modal perception system designed for indoor navigation and object interaction:
The combination of LiDAR for long-range spatial awareness and the RealSense camera for detailed close-range perception gives the G1 complementary sensing modalities. Researchers can also mount additional Intel RealSense D455 cameras as accessories for expanded perception capabilities [7].
The base G1 runs on an 8-core high-performance CPU that handles real-time motion control, sensor processing, and basic autonomy functions. The G1 EDU configurations add an NVIDIA Jetson Orin module providing up to 100 TOPS (tera operations per second) of AI computing power as a dedicated co-processor [27]. This enables on-device machine learning inference, computer vision pipelines, and more complex autonomous behaviors without relying on external computation.
Connectivity options include Wi-Fi 6 and Bluetooth 5.2 for wireless communication, plus a 5W speaker for audio output.
The G1 is powered by a 13-series smart lithium battery rated at 9,000 mAh, charged via a 54V 5A charger. Under typical mixed-use conditions (walking, standing, moderate sensor activity), the battery provides approximately two hours of operation, extending to three to four hours under standby or low-activity use [26]. Recharge time is approximately 1.5 to 2 hours with the included charger. The battery is designed for quick-swap replacement, allowing users to extend operating time by carrying spare packs.
The G1 walks at a maximum speed of approximately 2 m/s (7.2 km/h or 4.5 mph), making it one of the faster compact humanoid robots available [16]. While slower than the full-size H1 (which holds the Guinness World Record at 3.3 m/s), the G1's speed is notable given its smaller stature and lighter weight. The robot can navigate flat surfaces, climb stairs, and traverse moderately uneven terrain.
Locomotion is controlled through policies trained using reinforcement learning in simulation and transferred to hardware via sim-to-real transfer. Unitree provides the unitree_rl_gym framework on GitHub, which supports training custom locomotion policies for the G1 using NVIDIA Isaac Gym and MuJoCo simulation environments [8].
One of the G1's frequently demonstrated capabilities is its ability to recover from falls by standing up autonomously from both supine (face-up) and prone (face-down) positions. This self-righting behavior is trained using a two-stage reinforcement learning approach: the first stage discovers a viable getting-up trajectory under minimal constraints, while the second stage learns to robustly track that trajectory from varied starting poses and on different surfaces [9]. This capability is critical for real-world deployment, where falls are inevitable and manual intervention may not always be available.
The G1 has been showcased in a series of increasingly impressive dynamic demonstrations:
These demonstrations, while not directly representative of practical applications, serve as stress tests for the G1's control algorithms and actuator performance. The ability to execute high-dynamic maneuvers indicates robust underlying hardware and well-trained locomotion policies.
Beginning in mid-2025, Unitree used the G1 as the showcase platform for a series of public combat-sport and athletic events that tested locomotion, balance recovery, and remote-operator coordination under adversarial conditions.
These events, despite occasional public stumbles, demonstrated that the G1's hardware was durable enough to take repeated falls and impacts without catastrophic damage, a property prized in robotics research where hardware time is expensive.
The G1 EDU models ship with full software development kit support. Unitree provides the unitree_sdk2, which supports programming in C++, Python, and ROS 2. The SDK offers multiple levels of control:
ROS 2 compatibility allows researchers to integrate the G1 into the broader ROS ecosystem, using tools such as Gazebo for simulation, MoveIt for motion planning, and Nav2 for autonomous navigation. Unitree maintains open-source repositories on GitHub with URDF models, example code, and simulation configurations [8].
The base G1 (non-EDU) model does not include the full SDK and is primarily intended for demonstrations and pre-loaded behaviors rather than custom software development.
Unitree actively supports reinforcement learning research on the G1. The unitree_rl_gym repository provides training configurations for locomotion policies using frameworks such as NVIDIA Isaac Sim and Isaac Lab. Policies are trained with extensive domain randomization (varying friction, mass distributions, motor delays, and sensor noise) to produce robust behaviors that transfer from simulation to hardware without fine-tuning [8].
Research groups worldwide have published work on G1 locomotion, manipulation, and whole-body control using these tools. Over 30 research papers involving the G1 platform were published in 2025, spanning topics from gait-conditioned reinforcement learning to getting-up policies for humanoid robots [9].
The G1 has rapidly become one of the most-cited hardware platforms in the humanoid robotics literature. Notable open-source projects validated on the G1 include:
| Project | Institution / authors | Year | Contribution |
|---|---|---|---|
| HumanPlus | Stanford (Zipeng Fu, Qingqing Zhao, Qi Wu, Gordon Wetzstein, Chelsea Finn) | 2024 | Full-stack shadowing and imitation pipeline that lets the G1 follow human body and hand motion in real time from a single RGB camera, enabling teleoperated data collection for whole-body skills [28] |
| ASAP (Aligning Simulation and Real-World Physics) | Carnegie Mellon and collaborators | 2025 | Delta-action model trained with reinforcement learning on real-world data to reduce sim-to-real motion-tracking error on agile whole-body skills, validated end-to-end on the G1 [29] |
| Learning Getting-Up Policies | Multi-institution (arXiv:2502.12152) | 2025 | Two-stage reinforcement learning pipeline that lets the G1 stand up reliably from supine, prone, and tangled-limb starting poses on varied surfaces [9] |
| Gait-Conditioned RL with Multi-Phase Curriculum | arXiv:2505.20619 | 2025 | Curriculum learning approach for stable bipedal gaits on the G1 across walking, jogging, and trotting modes |
| Traversing Narrow Paths | arXiv:2508.20661 | 2025 | Two-stage RL framework for tightrope-style narrow-terrain locomotion on the G1 |
These papers, together with dozens of derivative works, established the G1 as the default real-world benchmark for new humanoid control algorithms. RoboUniversity, an academic program launched in 2025 with collaborators from Stanford and OpenMind, uses Unitree platforms (including the G1) for certified hands-on robotics curricula spanning beginner coding through advanced integration [21].
On January 29, 2026, Unitree open-sourced UnifoLM-VLA-0, a Vision-Language-Action (VLA) foundation model designed to serve as an "embodied brain" for the G1 [13]. Unitree describes it as "a Vision-Language-Action (VLA) large model in the UnifoLM series, designed for general-purpose humanoid robot manipulation" [30]. Built on top of Alibaba's Qwen2.5-VL-7B vision-language model, UnifoLM-VLA was continuously pretrained on a multi-task dataset and then trained on roughly 340 hours of real-world robot interaction data to develop physical intuition for manipulation tasks [13][30].
The model enables the G1 to autonomously perform household tasks through natural language commands. Unitree demonstrated that a single UnifoLM-VLA policy could reliably complete 12 categories of complex manipulation tasks, including opening and closing drawers, plugging and unplugging connectors, stacking blocks, and pick-and-place operations [13][30]. The full training code, inference pipeline, and model weights are available on GitHub and Hugging Face.
Unitree supports teleoperation of the G1 through extended reality (XR) devices, including Apple Vision Pro, PICO, and Meta Quest headsets. The teleoperation system includes integrated data recording, making it straightforward to collect demonstration datasets for imitation learning research. Unitree also open-sourced a teleoperation solution for humanoid robots, lowering the barrier for researchers who want to collect human demonstration data on the G1 platform.
The entry-level G1, priced from approximately $16,000 to $17,990, is a pre-configured platform with 23 DOF, basic grippers, and pre-loaded walking, dancing, and self-righting behaviors [22]. It does not include the full development SDK and is intended for demonstrations, exhibitions, human-robot interaction studies, and light task prototyping. The standard model includes an 8-month manufacturer warranty.
The EDU line targets researchers, educators, and developers who need full programmatic control of the robot. All EDU models include an NVIDIA Jetson Orin module (100 TOPS), complete ROS 2 and Python/C++ SDKs, technical documentation, and an 18-month warranty [27]. EDU configurations range from the U1 Standard ($43,900) through various Plus and Ultimate tiers up to approximately $73,900 for the most advanced configuration.
Key EDU tiers include:
| Model | Price (approx.) | Total DOF | Key additions |
|---|---|---|---|
| G1 EDU Standard (U1) | $43,900 | 23 | Jetson Orin, full SDK, ROS 2 |
| G1 EDU Plus (U2) | $53,900 | 29 | Extra waist and wrist DOF |
| G1 EDU Ultimate A (U3) | $65,900 | 43 | Dex3-1 three-finger hands |
| G1 EDU Ultimate D (U5) | $73,900 | 41 | Dex5-1 five-finger hands with up to 94 tactile sensors |
The G1 and H1 serve different roles within Unitree's product line. The H1 is a full-size humanoid platform aimed at industrial research, while the G1 is a compact, affordable alternative suited to academic labs, educational institutions, and developers entering humanoid robotics.
| Feature | Unitree G1 (Base) | Unitree G1 EDU Ultimate | Unitree H1 | Unitree H1-2 |
|---|---|---|---|---|
| Height | 132 cm | 132 cm | 180 cm | 178 cm |
| Weight | 35 kg | 35 kg | 47 kg | 70 kg |
| Total DOF | 23 | 43 | 19 | 27 |
| Max speed | ~2 m/s | ~2 m/s | 3.3 m/s | <2 m/s |
| Max knee torque | 90 N.m | 120 N.m | 360 N.m | 360 N.m |
| Dexterous hands | No | Yes (Dex3-1 or Dex5-1) | No | Optional |
| Jetson Orin | No | Yes (100 TOPS) | Optional | Optional |
| Price | ~$16,000 | ~$65,900-$73,900 | ~$90,000 | ~$128,900 |
| Primary audience | Demos, education | Research, development | Industrial research | Advanced research |
The G1's compact size, while advantageous for portability and safety, does limit its ability to work at standard human workbench heights. The H1's 180 cm stature allows it to interact with environments designed for adults, making it better suited for industrial and logistics applications.
The G1's primary market is academic and educational robotics. Its combination of affordability, open development tools, and ROS 2 compatibility makes it accessible to university labs that previously could not justify the cost of a humanoid platform. By 2026 industry observers described the G1 as the default research humanoid at "pretty much every university with a robotics program," with elite labs open-sourcing controllers that gave the robot what was widely characterized as superhuman whole-body capability for its size class [21]. Common research applications include:
Unitree has increasingly positioned the G1 for domestic applications. Demonstration videos have shown the G1 attempting cooking tasks, organizing household items, and performing basic tidying. With the release of UnifoLM-VLA in 2026, the robot gained the ability to respond to natural language commands and autonomously execute multi-step manipulation tasks in home environments [13].
However, practical household deployment remains limited. A widely shared viral video from early 2025 showed a G1 dressed in a maid outfit struggling with kitchen tasks in an unstructured home environment, illustrating the gap between controlled demonstrations and real-world domestic performance [14]. The robot excels in choreographed routines and structured research tasks but faces challenges with the unpredictability and variety of typical household environments.
The G1's compact form factor and low price have attracted interest from companies exploring humanoid robots for light assembly, palletization, quality inspection, and customer-facing service roles. While most current deployments are experimental, Unitree targets 10,000 to 20,000 humanoid shipments in 2026 as it scales production [5]. Total mass-production output across Unitree's humanoid lines exceeded 6,500 units by the end of 2025, a meaningful share of which were G1 variants destined for academic, commercial, and overseas customers [22].
The G1 occupies a distinctive position in the humanoid robot market as the most affordable production humanoid available as of 2026. The following table compares the G1 with other prominent humanoid platforms.
| Feature | Unitree G1 | Tesla Optimus | Boston Dynamics Atlas (Electric) | Figure 02 | Agility Digit |
|---|---|---|---|---|---|
| Height | 132 cm | 173 cm | 150 cm | 170 cm | 175 cm |
| Weight | 35 kg | 72 kg | 89 kg | ~60 kg | 65 kg |
| DOF (max config) | 43 | 28+ | 28+ | 24+ | 16+ |
| Max speed | ~2 m/s | ~1.8 m/s | ~2.5 m/s | N/A | 1.5 m/s |
| Price | From $16,000 | Not yet sold | Not for sale | Enterprise only | Lease only (~$250,000) |
| Publicly available | Yes | No (internal use) | No | No | Limited |
| SDK / ROS support | Yes | No public SDK | Limited | No | Limited |
| Backflip / acrobatics | Side flip | No | Yes (hydraulic version) | No | No |
The G1's core competitive advantage is availability. While Tesla Optimus, Boston Dynamics Atlas, and Figure 02 generate significant media attention and may ultimately prove more capable, none were commercially available to external buyers as of early 2026. Tesla uses Optimus internally in its Fremont factory, where units have largely been engaged in learning and data collection rather than productive manufacturing tasks, and the company has discussed potential external sales beginning in 2026 at an estimated $20,000 to $30,000 price point. Boston Dynamics' Atlas remains restricted to enterprise pilots with partners such as Hyundai. Figure AI's Figure 02 and the newer Figure 03 are in pilot deployment with industrial partners (including BMW) and are not sold to outside customers. Agility Robotics' Digit is available through lease arrangements but at a price point roughly fifteen times higher than the G1.
Unitree's strategy mirrors its approach in the quadruped market, where the company captured 60 to 70 percent of global shipments by offering capable robots at a fraction of competitors' prices. By making the G1 affordable enough for small labs, startups, and educational institutions, Unitree has positioned itself to build a large installed base and developer ecosystem, even if individual units are less capable than more expensive alternatives in certain dimensions. The contrast is especially stark with Atlas: published comparisons in 2025 placed the electric Atlas at roughly $420,000 versus the G1 at $16,000, a 26x price ratio for a platform with broadly similar degree-of-freedom count in its top configurations [23].
As Unitree's installed base grew through 2025, the G1 became the focal point of an emerging policy debate in the United States over Chinese-made robotic platforms. In September 2025, independent security researchers published findings that the G1's onboard software transmitted multi-modal sensor data to servers in China at five-minute intervals by default, without explicit notification in the user interface. The same disclosure described a wormable Bluetooth Low Energy (BLE) vulnerability that could allow an attacker in physical proximity to take full control of a G1, and identified a hard-coded static Blowfish-ECB encryption key shared across every G1 worldwide, which meant that compromising the key on a single robot would compromise the entire fleet [24].
In the months that followed, the bipartisan House Select Committee on the Chinese Communist Party sent a letter warning that Unitree's products, including the G1, posed a national-security risk because of the company's documented ties to People's Liberation Army research programs and military-civil-fusion initiatives. The committee noted that Unitree quadrupeds had already been deployed in U.S. state correctional and military facilities, and called for closer scrutiny of humanoid procurement [25]. A coalition of U.S. AI and robotics firms separately petitioned lawmakers for action, citing both competitive concerns and the data-collection findings.
As of mid-2026 no formal U.S. export or import ban had been enacted, and Unitree continued to ship G1 units to American buyers, including universities, research institutions, and private companies. Unitree has stated that newer firmware revisions allow users to disable cloud telemetry, but security researchers continue to recommend network segmentation when operating G1 hardware in sensitive environments.
Despite its strengths, the G1 has several notable limitations:
The G1 is manufactured at Unitree's facilities in Hangzhou, China. International orders typically ship within four to eight weeks, with some retailers quoting approximately two months for delivery. The robot ships with a battery, charger, and basic documentation. EDU models include the full SDK, ROS 2 integration package, and dedicated technical support.
Unitree reported total humanoid shipments exceeding 5,500 units in 2025, with mass-production output above 6,500 units across all humanoid lines, placing the company first globally in humanoid robot deliveries with about 32.4 percent of an estimated 17,000 units shipped worldwide [3][22]. Unitree posted 1.71 billion yuan in revenue for 2025, up roughly 335 percent year over year, with net profit rising nearly eightfold to about 600 million yuan; humanoid robots grew to 51.5 percent of revenue in the first nine months of 2025, up from 1.9 percent in 2023 [3]. Production targets for 2026 range from 10,000 to 20,000 humanoid units as Unitree scales manufacturing in anticipation of growing demand from research institutions, educational organizations, and commercial customers.