AgiBot X2
Last reviewed
Sources
20 citations
Review status
Source-backed
Revision
v5 · 4,119 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
20 citations
Review status
Source-backed
Revision
v5 · 4,119 words
Add missing citations, update stale details, or suggest a clearer explanation.
The AgiBot X2, also known as the Lingxi X2 (灵犀X2), is a compact bipedal humanoid robot developed by AgiBot (Zhiyuan Robotics), a Shanghai-based Chinese robotics company, and unveiled on March 11, 2025. Standing about 1.31 meters tall and weighing roughly 35 kilograms, the X2 is powered by AgiBot's GO-1 (Genie Operator-1) embodied foundation model and became, in September 2025, the first humanoid robot in the world to land a Webster flip, a forward somersault from a single-leg takeoff.[7][8]
The X2 is the successor to the open-source AgiBot X1 and represents the company's vision for a versatile, general-purpose embodied agent that combines motion intelligence, interaction intelligence, and task intelligence in a single platform. It is designed for applications in education, research, entertainment, security, hospitality, and light industrial assistance, and is significantly smaller than AgiBot's full-sized Yuanzheng A2 humanoid. The X2's GO-1 model enables real-time vision, speech, and gesture processing through the Vision-Language-Latent-Action (ViLLA) framework.[7][10]
AgiBot was the world's largest humanoid robot producer by shipment volume in 2025, delivering 5,168 units, about 39% of the global market, according to research firm Omdia, and is preparing a Hong Kong IPO targeting a valuation of up to roughly US$6.4 billion in 2026.[3][20]
AgiBot (Chinese: 智元机器人), officially AGIBOT Innovation (Shanghai) Technology Co., Ltd. and also known as Zhiyuan Robotics, was founded in February 2023 by Deng Taihua (CEO) and Peng Zhihui (CTO), both former Huawei engineers. Peng Zhihui, born in 1993 and known online as "Zhihui Jun," had risen to fame on Chinese video platform Bilibili for his DIY electronics projects, including an Iron Man-inspired robotic arm called "Dummy" built using Huawei's OpenHarmony operating system and Ascend Atlas AI processor. In 2020, Peng joined Huawei's "Top Minds" (also known as "Genius Youth") recruitment program, reportedly earning up to 3 million yuan (approximately US$420,000) annually. He departed Huawei in December 2022 to co-found AgiBot.[1][2]
Within just two years, AgiBot grew from a startup to the world's leading humanoid robot shipper. Research firm Omdia ranked AgiBot No. 1 worldwide in 2025 humanoid robot shipments, with 5,168 units delivered, an estimated 39% of the roughly 13,000 humanoid robots shipped globally that year; the company produced its 5,000th robot in December 2025 and its 10,000th by March 2026.[15][20] AgiBot's investors include BYD, Tencent, HongShan Capital (formerly Sequoia China), Hillhouse Investment, LG Electronics, and Mirae Asset. According to Reuters, AgiBot is planning a Hong Kong IPO with a target valuation of HK$40 billion to HK$50 billion (about US$5.1 billion to US$6.4 billion), planning to file a preliminary prospectus in early 2026 with a listing targeted by the third quarter of 2026; CICC, CITIC Securities, and Morgan Stanley are among the advising banks. Reuters reported that "AgiBot plans to file a preliminary prospectus early next year ... with one [source] adding that the firm aims for a public listing by the third quarter of 2026," though the offering's size and valuation remain subject to change.[3][4]
The X2 belongs to AgiBot's Lingxi (灵犀) series, a line of compact bipedal humanoids designed to be roughly half the height of an adult human. The series is distinct from AgiBot's full-sized Yuanzheng (远征, "Expedition") A-series, which targets commercial and industrial deployments, and the Genie G-series of wheeled industrial robots.
The first robot in the Lingxi line was the AgiBot X1, released in 2024 as a fully open-source research platform. The X1 was designed to fill a gap in the humanoid market by providing researchers, universities, and developer communities with a hackable bipedal platform that included open hardware documentation, full source code, and the AimRT robotics middleware framework. While the X1 prioritized openness and modularity, the X2 shifted focus toward commercial viability, advanced AI-driven interaction, and dynamic physical performance.[5]
The AgiBot X2 is a compact humanoid robot with the following core specifications:
| Specification | AgiBot X2 (Standard) | AgiBot X2 Ultra |
|---|---|---|
| Height | 1,310 mm (4 ft 4 in) | 1,310 mm (4 ft 4 in) |
| Width | 460 mm | 460 mm |
| Depth | 210 mm | 210 mm |
| Weight | ~35 kg (77 lb) | ~39 kg (86 lb) |
| Total Degrees of Freedom | 25 | 30-31 |
| Arm DOF (per arm) | 5 | 7 |
| Leg DOF (per leg) | 6 | 6 |
| Waist DOF | 3 | 3 |
| Neck DOF | 0 | 1-2 |
| Arm Reach | 558 mm (excl. end effector) | 558 mm (excl. end effector) |
| Peak Joint Torque | 120 N-m | 120 N-m |
| Walking Speed (max) | 1.8 m/s | 1.8 m/s |
| Walking Speed (typical) | 0.8 m/s or less | 0.8 m/s or less |
| Payload (full range) | 1 kg or less | 1 kg or less |
| Payload (specific postures) | Up to 3 kg | Up to 3 kg |
| Battery Capacity | ~500 Wh | ~500 Wh |
| Runtime | ~2 hours at 0.5 m/s | ~2 hours at 0.5 m/s |
| Charging Time | 1.5 hours or less | 1.5 hours or less |
| Main Compute | RK3588 x 2 | RK3588 x 2 + Orin NX (157 TOPS) |
| Operating Temperature | -10 to 40 degrees C | -10 to 40 degrees C |
| Auto-Charging Dock | No | Yes (optional) |
| 3D LiDAR | No | Yes |
| RGB-D Camera | No | Yes |
| Secondary Development | Not supported | Supported |
| OmniHand Compatibility | No | Yes |
These figures are drawn from AgiBot's official X2 product page. Some early pre-launch coverage cited slightly different numbers (for example, 33.8 kg and 28 degrees of freedom); the manufacturer's published specifications above are the authoritative reference.[6]
The X2 integrates several proprietary hardware subsystems developed in-house by AgiBot:
The standard X2 is equipped with an interactive RGB camera capable of facial recognition, object detection, and gesture recognition, along with a head-mounted touch sensor, a microphone array, and speakers. The X2 Ultra variant adds significantly more sensing capability: a 3D LiDAR unit for point-cloud-based SLAM (Simultaneous Localization and Mapping), an RGB-D depth camera for 3D spatial perception, front-facing binocular RGB cameras, a rear-facing RGB camera, and 4G/5G connectivity.
Both models feature an interactive display screen with lighting effects for expressive communication, as well as Wi-Fi and Bluetooth connectivity.
Both X2 variants use dual Rockchip RK3588 processors as their main compute units. The X2 Ultra supplements this with an NVIDIA Jetson Orin NX module delivering 157 TOPS (trillion operations per second) of AI computing performance, enabling on-device execution of more sophisticated deep learning models for autonomous navigation, obstacle avoidance, and real-time inference. The Ultra variant also supports secondary development, allowing researchers and integrators to deploy custom AI models and applications.
The X2 represents a significant evolution from the X1 in terms of design philosophy, AI integration, and target use case.
| Feature | AgiBot X1 | AgiBot X2 |
|---|---|---|
| Height | 130 cm | 131 cm |
| Weight | 33 kg | 35 kg |
| Degrees of Freedom | 34 | 25 (standard), 30-31 (Ultra) |
| Walking Speed | 3.6 km/h (1.0 m/s) | 6.5 km/h (1.8 m/s, max) |
| Battery Runtime | ~2 hours | ~2 hours |
| Arm Payload | 0.5 kg per arm | Up to 3 kg (specific postures) |
| Peak Joint Torque | 200 N-m (R86-3 actuator) | 120 N-m |
| Servo Technology | PowerFlow (open-source docs) | PowerFlow + Xyber-Edge |
| AI Foundation Model | Not integrated | GO-1 (ViLLA framework) |
| Open Source | Fully open (hardware, firmware, middleware) | Partially open (Ultra supports secondary dev) |
| Gripper | OmniPicker adaptive gripper | OmniHand (Ultra only) |
| Autonomous Navigation | Basic | Advanced (Ultra with LiDAR + RGB-D SLAM) |
| Target Use Case | Research, education, developer community | Commercial, entertainment, research, service |
| Price | ~US$20,000 | From ~US$27,300 |
The X1 offers more degrees of freedom (34 vs. 25-31) because its design prioritized giving researchers maximum joint-level control. The X2 trades some joint count for higher walking speed, improved payload capacity, deeper AI integration through GO-1, and a more polished commercial package with expressive interaction capabilities. The X1 remains the recommended platform for robotics research requiring full hardware and software transparency, while the X2 is better suited for deployment scenarios involving human-robot interaction, demonstrations, and light service tasks.[5][6]
AgiBot markets the X2 around three pillars of intelligence: motion intelligence, interaction intelligence, and task intelligence.
The X2 demonstrates exceptional locomotion agility for its size class. Beyond standard walking and running, the robot can perform complex dance routines, ride a scooter, balance on a hoverboard, and pedal a bicycle. AgiBot claims these capabilities "far surpass similar products in motion flexibility."[7]
The most dramatic demonstration of the X2's motion intelligence came in September 2025, when AgiBot released a slow-motion 4K video showing the X2 performing a Webster flip: a forward somersault initiated from a single-leg, back-facing takeoff. This acrobatic maneuver demands asymmetric explosive power, precise midair body control, and accurate landing stabilization. The X2 became the first humanoid robot in the world to execute this move, and the video went viral across global social media and robotics communities. AgiBot humorously captioned the video: "You may not know me, but you must have heard that real men can do the Webster flip," referencing a popular internet meme.[8][9]
These dynamic capabilities are enabled by AgiBot's proprietary motion control algorithms running on the Xyber-Edge cerebellum controller and PowerFlow joint actuators, which provide the torque density and response speed needed for high-energy maneuvers.
AgiBot describes the X2 as the "first truly interactive dynamic robot." The X2 employs multimodal interaction with millisecond-level response times: it can detect human emotional states through facial expression analysis and voice tone recognition, and responds with lifelike characteristics including mimicked breathing rhythms, subtle body language movements, head nods, and expressive gestures.[7]
The interaction system processes inputs from the robot's cameras, microphone array, and touch sensors to build a real-time model of the humans around it. This enables the X2 to conduct natural conversations, react to social cues, and adapt its behavior to context, making it suitable for reception, guided tours, entertainment performances, and educational demonstrations.
The X2's task intelligence is built on AgiBot's GO-1 foundation model, which provides zero-shot generalization capabilities for object manipulation in simple tasks. The robot can process image and video data to understand its environment, plan actions, and execute tasks without requiring pre-programmed routines for every scenario. The GO-1 model also supports multi-robot collaboration, allowing multiple X2 units (or mixed fleets of AgiBot robots) to coordinate on complex assignments such as security patrol or event logistics.[7]
Primary application domains for the X2's task intelligence include security monitoring, housekeeping, light cleaning, inventory checks, and guided demonstrations across education and healthcare facilities.
The X2 is powered by AgiBot's Genie Operator-1 (GO-1), a generalist embodied foundation model released on March 10, 2025. GO-1 introduces the Vision-Language-Latent-Action (ViLLA) framework, which represents an evolution beyond the standard Vision-Language-Action (VLA) paradigm used in prior robot learning models, in which, as AgiBot describes it, "actions are directly conditioned on vision and language inputs."[10]
The ViLLA framework combines a Vision-Language Model (VLM) with a Mixture of Experts (MoE) architecture consisting of two key components:
In benchmark evaluations, GO-1 increased task success rates by 32 percentage points over prior state-of-the-art models (from 46% to 78%), with the Latent Planner alone contributing a 12-percentage-point improvement (from 66% to 78%). The model achieves over 60% success rate on complex, long-horizon dexterous manipulation tasks.[10][11]
The training data underlying the X2's AI capabilities comes from AgiBot World Colosseo, one of the largest open-source robot manipulation datasets ever released. The dataset was a finalist for the IROS 2025 Best Paper Award and was published in IEEE Transactions on Robotics (TRO) in 2026.[12]
| Metric | Value |
|---|---|
| Total Trajectories | 1,003,672 (~43.8 TB) |
| Curated Alpha Subset | 92,214 trajectories (~8.5 TB) |
| Number of Tasks | 217 |
| Number of Skills | 87 |
| Number of Scenes | 106 |
| Total Duration | 2,976.4 hours |
| Number of Objects | 3,000+ |
| Deployment Domains | 5 (domestic, retail, industrial, restaurant, office) |
| Collection Facility | 4,000+ sq. meters, Shanghai |
| License | CC BY-NC-SA 4.0 |
The data was collected by over 100 homogeneous robots operated by skilled teleoperators in a purpose-built facility that replicates real-world environments. Extensive human annotations are provided at the item, scene, skill (sub-task segmented), and task levels. The dataset also includes imperfect data with annotated error states to help train models that can recover from failures, along with tasks requiring dexterous hand manipulation.[12][13]
Policies pre-trained on AgiBot World achieve an average success rate improvement of 30% compared to those trained on the Open X-Embodiment (OXE) dataset, which had previously been the largest open robot learning dataset.[12]
The X2 is offered in multiple configurations to serve different market segments.
The entry-level variant emphasizes cost-effectiveness with a reduced degree-of-freedom count (approximately 27 DOF). It targets price-sensitive buyers in education and basic demonstration roles.
The Pro model increases dexterity to approximately 31 DOF, including 7-DOF arms that enable more humanlike reach and in-hand manipulation. It is positioned as a balance between capability and cost, with configurable sensor options and support for open-stack development workflows suitable for academic and R&D use.
The top-tier variant adds the NVIDIA Orin NX computing module (157 TOPS), 3D LiDAR, RGB-D camera, front binocular and rear RGB cameras, 4G/5G connectivity, and compatibility with AgiBot's OmniHand dexterous manipulation system. A key differentiator is the optional auto-charging dock, which allows the Ultra to return to its station and recharge autonomously, enabling continuous duty cycles for exhibitions, waypoint tours, or concierge scenarios without manual battery swaps. The Ultra also supports secondary development for custom AI model deployment.
The X2-N is a specialized variant featuring a unique hybrid locomotion system with integrated foot wheels, allowing it to transition seamlessly between bipedal walking and wheeled rolling. In bipedal mode, it can climb stairs and traverse uneven terrain; in wheeled mode, it can move rapidly across flat surfaces, ramps, and narrow bridges. The X2-N can carry loads of up to 5.4 kg while walking.[14]
A notable feature of the X2-N is its sensor-less navigation approach. Unlike other humanoid robots that rely heavily on cameras and external sensors, the X2-N uses proprioceptive feedback from joint torque sensors, pressure sensors, and internal gyroscopes to sense and adapt to terrain in real time, mimicking human intuition about ground conditions. This approach reduces hardware costs and computational overhead while improving robustness in environments where visual sensors might be unreliable (dusty factories, low-light conditions).[14]
The X2-N is targeted at logistics, search-and-rescue, and industrial automation applications where versatile mobility across mixed terrain is essential.
While the X2 belongs to AgiBot's compact Lingxi line, the company's primary commercial humanoid is the Yuanzheng A2 series, a full-sized bipedal robot standing 169-175 cm tall and weighing 55-69 kg.
| Feature | AgiBot X2 | AgiBot A2 |
|---|---|---|
| Series Name | Lingxi (灵犀) | Yuanzheng (远征, Expedition) |
| Height | 131 cm | 169-175 cm |
| Weight | 35 kg | 55-69 kg |
| Degrees of Freedom | 25-31 | 40-49+ |
| AI Computing | Up to 157 TOPS (Ultra) | 200 TOPS |
| Arm Payload | Up to 3 kg | 3-5 kg per arm (up to 15 kg) |
| Battery | ~500 Wh | 700 Wh (standard); 2,000 Wh (A2-W) |
| Walking Speed | Up to 1.8 m/s | Up to 3.3 m/s |
| Primary Market | Education, entertainment, research, light service | Industrial automation, commercial service, manufacturing |
| Variants | Lite, Pro, Ultra, X2-N | Standard, Max, Ultra, Lite, A2-W (wheeled) |
The A2 series has been deployed across more than 10 industries and was the primary driver behind AgiBot's market-leading shipment numbers in 2025. In November 2025, an A2 set a Guinness World Record by walking 106.286 kilometers from Suzhou to Shanghai over three days. The A2 also won multiple Best of CES 2026 awards at AgiBot's U.S. market debut.[15][16]
The X2 and A2 are complementary products. The X2's smaller form factor, lower cost, and expressive interaction capabilities make it ideal for environments where a full-sized humanoid would be impractical or intimidating, such as schools, museums, retail showrooms, and entertainment venues. The A2 is the workhorse for factory floors, warehouse logistics, and professional service applications demanding greater payload capacity and endurance.
AgiBot positions the X2 across eight core commercial application categories:
At CES 2026, AgiBot showcased the X2 as part of its full product portfolio during the company's official U.S. market debut. The X2 was demonstrated alongside the A2 Series, G2 industrial robot, D1 quadruped, and OmniHand dexterous manipulator. AgiBot took home multiple Best of CES 2026 awards, including Best of Show from Ubergizmo.[16]
The X2 is also available for lease through AgiBot's BotShare platform (co-launched in December 2025), which operates across 50 cities in China with over 600 service providers. Rental rates start at approximately 899 euros per day through AgiBot's European distribution partner.
In July 2025, AgiBot released Lingqu OS, described as the world's first embodied intelligent operating system. Lingqu OS provides a layered, open-source robotics software stack consisting of:
Lingqu OS supports over-the-air (OTA) updates and mobile app control for the X2, allowing operators to push new capabilities and behaviors to deployed robots without physical access.[17]
Genie Sim 3.0 is AgiBot's simulation platform for training and validating robot policies in virtual environments before deploying them to physical hardware. It won a Best of Show 2026 award from Ubergizmo at CES for its role in connecting AgiBot's entire robotics portfolio through a unified software platform.[16]
Announced in October 2025, LinkCraft is a zero-code platform that converts human motion videos into executable robot action sequences. This allows non-technical users to "teach" the X2 new behaviors simply by demonstrating the desired motion on video, significantly lowering the barrier to programming new tasks.[18]
The AgiBot X2 competes in the growing market for compact humanoid robots. The broader humanoid robot industry saw Chinese companies account for approximately 90% of global shipments in 2025, with AgiBot and Unitree Robotics together shipping more than 10,000 units, far ahead of Western competitors such as Tesla Optimus, Figure AI, and Agility Robotics, each of which shipped roughly 150 units.[19]
In the compact humanoid segment specifically, the X2 competes with Unitree's G1 (priced from US$13,500), which also targets education and research markets and offers configurations ranging from 23 to 43 degrees of freedom. AgiBot differentiates the X2 through its deeper AI integration via the GO-1 foundation model, its expressive interaction capabilities, and its broader ecosystem of software tools (Lingqu OS, AgiBot World, Genie Sim). Tesla's Optimus, while more advanced in some respects, is a full-sized humanoid targeting a different market segment and price point (US$20,000-$30,000 target).
The X2's pricing starts at approximately US$27,300 for the standard model, positioning it above the cheapest Chinese alternatives but well below the six-figure price points of Western industrial humanoids from companies like Boston Dynamics or Agility Robotics.