CtrlK
| XPENG Robotics | |
|---|---|
| General information | |
| Full name | Shenzhen Pengxing Intelligence Co., Ltd. |
| Also known as | Pengxing Intelligence, XPENG Robotics |
| Parent company | XPeng Inc. (NYSE: XPEV) |
| Founded | December 22, 2020 |
| Founder | He Xiaopeng |
| Headquarters | Shenzhen, Guangdong, China |
| Key people | Mi Liangchuan (VP, Robotics Division); Chen Jie (RL Lead); Ge Yixiao (Multimodal AI); Liu Xianming (Autonomous Driving) |
| Industry | Robotics, Artificial intelligence |
| Products | XPENG IRON, XPENG PX5, Xiaobailong (robot horse) |
| Employees | ~200 (2025) |
| Website | pxing.com |
XPENG Robotics, formally known as Shenzhen Pengxing Intelligence Co., Ltd. (Chinese: Pengxing Zhineng), is the robotics division of Chinese electric vehicle manufacturer XPeng Inc. (NYSE: XPEV). Founded in December 2020 following XPeng's acquisition of quadruped robot startup Dogotix, the division develops humanoid robots, quadruped platforms, and related artificial intelligence systems. XPENG Robotics is headquartered in Shenzhen, Guangdong, with a dedicated robot manufacturing facility under construction in Guangzhou.[1][2]
The division represents what XPeng chairman He Xiaopeng has called the company's "third growth curve" after electric vehicles and global expansion. Its strategic thesis centers on the idea that autonomous driving is the "smallest implementation unit of embodied AI," and that the perception, planning, and control technologies developed for self-driving cars can be systematically transferred to humanoid robotics. Approximately 70% of the technology stack powering XPENG Robotics' products is shared with XPeng's automotive division.[3]
XPENG Robotics has progressed through multiple robot generations since its founding, beginning with quadruped platforms and advancing to bipedal humanoid designs. Its key products include the XPENG PX5 prototype (2023), the first-generation XPENG IRON (2024), and the next-generation IRON (2025). The company plans to begin mass production of humanoid robots by the end of 2026 and has announced a long-term investment plan of approximately 100 billion yuan ($13.8 billion USD) for its robotics program, with a target of selling over one million units annually by 2030.[4][5]
XPeng's involvement in robotics can be traced to the personal interest of its founder and chairman, He Xiaopeng. He began conducting serious research into robot technologies as early as 2016, meeting regularly with prominent figures in the Chinese robotics industry. Among those he held early conversations with was Wang Xingxing, the founder of Unitree Robotics.[6]
He Xiaopeng came to view robotics as a natural extension of XPeng's core autonomous driving business. His reasoning centered on the observation that the technology chain of "perception, decision, and execution" in the automotive field is highly analogous to that required for robots. Intelligent driving, in his framing, represents the smallest implementation unit of embodied intelligence. This insight became foundational to XPENG Robotics' approach of transferring autonomous driving technology directly to its robot platforms.[3]
In the autumn of 2020, He Xiaopeng traveled to Shenzhen to personally negotiate the acquisition of Dogotix, a startup company specializing in quadruped robot research and development. Dogotix was founded by Zhao Tongyang and was recognized as one of the earliest teams in China to achieve the practical implementation of quadruped robot technology. Under Zhao's leadership, the team had completed the R&D of its first-generation robot dog in fewer than two months.[6]
Xiaomi founder Lei Jun reportedly helped facilitate the deal. XPeng spent approximately $100 million (USD) to buy out the shares of Dogotix's existing investors, bringing the company fully under XPeng's corporate umbrella. The acquisition gave XPeng immediate access to expertise in legged locomotion, dynamic balance control, and robot actuation systems.[7]
Following the Dogotix acquisition, XPeng formally established Shenzhen Pengxing Intelligence Co., Ltd. on December 22, 2020, with a registered capital of 10 million yuan. The subsidiary was headquartered at Shenzhen Bay Science and Technology Ecological Park and served as XPeng's dedicated robotics division.[1]
The core engineering group, however, had deeper roots. The team was originally established in Shenzhen in 2016 and was one of the earliest Chinese teams to invest in the research and development of legged robots. After being absorbed into XPeng's ecosystem, Pengxing was positioned as part of the company's "Future Transportation Explorer" strategy, leveraging automotive systems and concepts to empower the R&D and manufacturing of intelligent robots.[1]
At its peak in the early years, the Pengxing team grew to more than 300 employees. However, internal misalignments, including strategic disagreements between Zhao Tongyang and He Xiaopeng, eventually led to the departure of core team members. Zhao left XPeng and went on to found EngineAI (also known as Zhongqing Robotics), a competing humanoid robot company. After these departures, the team was reduced to approximately 70 people before being rebuilt under new leadership.[6]
In 2023, He Xiaopeng appointed Mi Liangchuan to lead the robot business. Mi had joined XPeng in 2021 as senior director of autonomous driving, where he led the team in building the algorithm and architecture for XPeng's assisted driving system. Before XPeng, Mi spent 15 years at NVIDIA, rising from software engineer to senior Android software manager at NVIDIA's Santa Clara headquarters, where he managed a team of hundreds of people and reported to NVIDIA's global vice-president Wu Xinzhou. He holds degrees from the University of Science and Technology of China (electrical engineering) and Iowa State University (robotics).[8][9]
Under Mi's leadership, the team stabilized and began scaling back up, reaching approximately 200 employees by 2025.[6]
XPENG Robotics has developed a total of at least eight generations of robots, progressing from quadruped designs to increasingly sophisticated bipedal humanoid platforms.
Before pivoting to humanoid form factors, Pengxing Intelligence produced several quadruped robot platforms. The most notable was the Xiaobailong (literally "Little White Dragon"), unveiled in September 2021 as the world's first rideable intelligent robot horse. The third-generation prototype featured a car-level intelligent driving system incorporating cameras and LiDAR for environment perception, map building, route planning, autonomous movement, target following, and automatic obstacle avoidance. The robot horse also included voice interaction, face and voiceprint recognition, emotional expression through a face screen, and tactile perception.[10]
Following the Xiaobailong, the team developed a children's robot pony and a more advanced "Unicorn" variant. The Unicorn featured a software system based on XPeng's autonomous vehicle architecture and a robotic tail with six degrees of freedom based on XPENG Robotics' X-arm 2.0 platform.[11]
XPeng developed five quadruped designs in total before concluding that four-legged robots lacked hands for manipulation tasks and had difficulty navigating complex home and workplace environments. This assessment prompted the pivotal shift to bipedal humanoid designs.[12]
| Product | Year | Type | Key features |
|---|---|---|---|
| Xiaobailong (robot horse) | 2021 | Quadruped | World's first rideable robot horse; car-level intelligent driving; LiDAR + camera perception |
| Robot pony / Unicorn | 2022 | Quadruped | Children's companion; autonomous vehicle software architecture; X-arm 2.0 tail |
| XPENG PX5 | 2023 | Bipedal humanoid | 150 cm, 109 kg; 22 DOF; 430 g hands; 0.05 mm arm accuracy |
| XPENG IRON (1st gen) | 2024 | Bipedal humanoid | 178 cm, 70 kg; 60+ joints; 1x Turing AI chip; VLA 1.0 |
| XPENG IRON (next-gen) | 2025 | Bipedal humanoid | 82 DOF; 22 DOF per hand; 3x Turing chips (2,250 TOPS); VLA 2.0; all-solid-state battery |
| ET1 | 2026 | Bipedal humanoid | First prototype built to automotive-grade manufacturing standards |
The XPENG PX5 was publicly unveiled on October 24, 2023, at XPeng's annual 1024 Tech Day event in Guangzhou. Standing 150 cm tall and weighing 109 kg, the PX5 was XPENG Robotics' first humanoid robot and represented the company's transition from quadruped to bipedal platforms. It featured 22 degrees of freedom, self-developed high-performance joints, and ultra-lightweight dexterous hands weighing just 430 grams each with 11 degrees of freedom per hand.[13]
The PX5 achieved stable walking capabilities after just five months of R&D, a pace attributed to the team's ability to leverage existing autonomous driving algorithms and sensor fusion technologies from XPeng's EV business. During its public demonstration, the robot navigated grass, rocky surfaces, and indoor environments, played football, and rode a self-balancing scooter while maintaining stability even when kicked. The PX5 validated XPeng's core thesis that automotive AI technology could transfer effectively to humanoid robotics.[14]
The XPENG IRON was unveiled on November 6, 2024, at XPeng's AI Day in Guangzhou, representing a dramatic leap forward from the PX5. Standing 178 cm tall and weighing just 70 kg (compared to the PX5's 109 kg at 150 cm), IRON featured more than 60 active joints, a 1:1 human-scale hand design with 15 degrees of freedom per hand, and a 20% improvement in torque density over the PX5's joint actuators. It was powered by XPeng's self-developed Turing AI chip.[15]
The next-generation IRON, unveiled at XPeng's 2025 AI Day on November 5, 2025, pushed the platform further: 82 degrees of freedom, 22 degrees of freedom per hand with the industry's smallest harmonic joints, three Turing AI chips delivering a combined 2,250 TOPS of computing power, and the world's first all-solid-state battery in a humanoid robot. This model adopted an "extreme anthropomorphism" design philosophy featuring a biomimetic spine, bionic muscles, flexible synthetic skin with embedded touch sensors, and a 3D curved OLED display face. Its movements proved so lifelike that He Xiaopeng had staff physically cut open the robot's leg on stage to prove no human was hidden inside.[16][17]
In January 2026, XPeng announced the completion of its first "ET1" prototype robot, developed to automotive-grade manufacturing standards. CEO He Xiaopeng wrote on Weibo: "The first unit of our ET1 robot, developed to automotive-grade standards, has successfully rolled off the production line." The ET1 milestone demonstrated the successful integration of XPeng's vehicle engineering and quality assurance expertise with its robotics capabilities, marking a key step toward mass production readiness.[18]
The Turing chip is XPeng's self-developed AI processor, used across both its automotive and robotics products. Each chip features a 40-core processor and delivers 750 TOPS of effective computing power. The chip can support AI models with up to 30 billion parameters, enabling on-device inference without reliance on cloud computing for core decision-making.[15][19]
In the next-generation IRON, three Turing chips work in parallel to provide 2,250 TOPS of total processing capability. XPeng claims this figure is three times higher than the second-best humanoid robot on the market and ten times higher than mainstream industry benchmarks. The same Turing chip architecture also powers XPeng's Robotaxi vehicles, where four chips deliver up to 3,000 TOPS. The chip completed over 2,700 functional verifications in just 40 days during development, achieving three times the industry standard for development efficiency.[19][20]
| Turing chip application | Number of chips | Total computing power |
|---|---|---|
| XPENG IRON (1st gen, 2024) | 1 | 750 TOPS |
| XPENG IRON (next-gen, 2025) | 3 | 2,250 TOPS |
| XPeng Robotaxi | 4 | 3,000 TOPS |
IRON's intelligence is built on a tri-model AI architecture combining three interconnected large models, each adapted from XPeng's autonomous driving technology stack:
| Model | Full name | Role |
|---|---|---|
| VLT | Vision-Language-Task | Core decision-making engine; enables autonomous reasoning and multi-step task planning in the physical world |
| VLA | Vision-Language-Action | Action generation model; translates sensory input into motor commands across all 82 joints |
| VLM | Vision-Language-Model | Multimodal perception and natural language interaction |
The VLT (Vision-Language-Task) model, first announced at the 2025 AI Day, serves as IRON's "brain" for autonomous decision-making. The VLA (Vision-Language-Action) model, now in its second generation, was adapted directly from XPeng's autonomous driving division and handles the translation of visual information into physical movements. The VLA 2.0 model was trained on nearly 100 million video clips. The VLM (Vision-Language-Model) provides multimodal understanding and conversational capabilities. Working together, the three models enable IRON to simultaneously converse, walk, and interact with its environment.[16][21]
XPeng refers to this overall framework as its first-generation Physical-World Large Model.
IRON runs XPeng's Tianji AIOS (AI Operating System), a real-time software platform designed for full-body coordination, balance control, task planning, and environmental awareness. Tianji AIOS manages the integration of perception, control, and voice interaction systems across the robot's sensor suite and actuator network.[22]
The Eagle-Eye perception system provides a 720-degree field of view, encompassing full 360-degree horizontal and 360-degree vertical coverage. Adapted from XPeng's autonomous driving sensor suite (known as Hawkeye in its automotive application), the system uses a combination of RGB cameras, 3D depth cameras, and LiDAR to build real-time three-dimensional maps of the robot's surroundings through SLAM (Simultaneous Localization and Mapping) algorithms. XPeng states the system can identify objects, gauge distances, and even read micro-expressions on human faces for social interaction.[23]
The next-generation IRON uses an all-solid-state battery, which XPeng describes as the first such application in any humanoid robot. Traditional lithium-ion batteries rely on flammable liquid electrolytes, which present safety risks when robots operate near people. Solid-state batteries replace these liquids with ceramic or polymer electrolytes, improving safety while offering higher energy density. XPeng states that IRON's solid-state battery achieves an energy density exceeding 500 Wh/kg, roughly double the capacity of the battery used in Tesla Optimus within the same physical footprint.[24]
XPeng's AI training infrastructure supports both its automotive and robotics divisions. The company has partnered with Alibaba Cloud since 2022 to build a large AI-native data center. XPeng's total computing power reaches 2.51 EFLOPS, dedicated to training its AI models. The company's dataset comprises 20 million data clips, and it iterates its models every two days.[25]
To address the humanoid robotics industry's persistent challenge of insufficient high-quality training data, XPENG Robotics has established its first embodied intelligence data factory in Guangzhou. This facility is dedicated to collecting behavioral data from robot-environment interactions, enabling the company's AI models to continuously improve through extensive physical training sessions.[16]
A defining characteristic of XPENG Robotics' strategy is the extensive reuse of technology originally developed for XPeng's electric vehicles. He Xiaopeng has stated that approximately 70% of IRON's technology stack derives from the company's EV platform.
| Technology area | EV origin | Robot application |
|---|---|---|
| AI chip | Turing chip for ADAS/autonomous driving | Core computing for VLT + VLA + VLM models |
| Vision system | Hawkeye/Eagle-Eye for XNGP driving | 720-degree spatial perception |
| VLA model | End-to-end autonomous driving model | Action generation from visual input |
| E/E architecture | XEEA vehicle electrical/electronic systems | Robot control and communication backbone |
| Neural networks | XNet 2.0 BEV perception | Spatial awareness and obstacle detection |
| Cloud AI | Shared cloud infrastructure for fleet learning | Robot training data processing |
| Control systems | Vehicle dynamics and stability control | Bipedal balance and locomotion |
This convergence strategy allows XPeng to amortize R&D costs across its vehicle and robotics businesses while accelerating the development of both product lines. The shared technology base contributed to the PX5's five-month development timeline from project start to stable bipedal walking, and continues to accelerate IRON's path to mass production. He Xiaopeng has argued that "humanoid robotics and autonomous driving can accelerate one another when built on shared infrastructure."[3][26]
XPENG Robotics is led by a team of senior engineers with backgrounds spanning NVIDIA, autonomous driving, and AI research.
| Name | Role | Background |
|---|---|---|
| Mi Liangchuan | VP, Robotics Division; Head of AI Technology Committee | 15 years at NVIDIA; senior Android software manager; University of Science and Technology of China + Iowa State University |
| Chen Jie | Reinforcement Learning Lead (joined July 2025) | University of Hong Kong (undergrad); Harvard University (doctorate); former ByteDance Seed team + Momenta |
| Ge Yixiao | Founding Director, Intelligent Mimetic Department | Huazhong University of Science and Technology; doctorate from Chinese University of Hong Kong MMLab; former Tencent ARC Lab chief researcher; 10,000+ Google Scholar citations |
| Liu Xianming | Autonomous Driving Integration Lead (joined March 2024) | Dual doctorates from Harbin Institute of Technology + University of Illinois; former Microsoft Research Asia, Baidu, Adobe, Facebook, Cruise |
Mi Liangchuan is considered the "number one person" of XPeng's robot project, responsible for coordinating the technical route, organizational structure, and product implementation. He is involved in almost every key link of the humanoid robot initiative.[8][9]
In February 2026, XPENG announced plans to break ground on a dedicated humanoid robot mass production facility at the Guangtang Sci-Tech Innovation City Embodied Intelligence Industrial Park in Guangzhou's Tianhe District. The 110,000-square-meter facility will function as the industry's first "full-chain" manufacturing hub, covering processes from R&D validation and small-batch trial production to large-scale manufacturing. Phase one construction includes robot factory buildings, power stations, and supporting infrastructure.[27][28]
Mass production preparation is scheduled to begin in April 2026, with the goal of achieving large-scale manufacturing by the end of 2026. The units intended for the late 2026 rollout will be "eighth generation" models. If XPENG meets this timeline, IRON could become one of the first mass-produced next-generation humanoid robots, potentially reaching the market ahead of competitors such as Tesla's Optimus.[16][27]
XPeng has committed up to 100 billion yuan (approximately $13.8 billion USD) in long-term investment over the next two decades to drive humanoid robotics development and scale IRON's production. He Xiaopeng has outlined a target of selling more than one million humanoid robot units annually by 2030. On pricing, He noted that "software accounts for over 50% of a robot's value from day one, compared to just 10-20% in traditional vehicles," and stated the company aims for retail pricing "very similar to car prices," potentially making robots accessible to households within five years.[4][5]
| Milestone | Target date |
|---|---|
| ET1 prototype (automotive-grade) | January 2026 (completed) |
| Factory groundbreaking (Guangzhou) | Q1 2026 |
| Mass production preparation | April 2026 |
| Large-scale manufacturing | End of 2026 |
| One million units annually | 2030 |
Hundreds of IRON units have already been deployed on XPeng's manufacturing lines in Guangzhou, where they handle sorting, transport, and screw installation operations on the XPENG P7+ production line. The company reports that a single IRON unit can install up to 500 screws per day, translating to annual labor cost savings exceeding 300,000 yuan per robot.[29]
XPENG Robotics' initial commercialization roadmap prioritizes controlled commercial environments rather than domestic or heavy industrial settings. The first deployment wave will position IRON units in people-facing roles:
| Application | Description |
|---|---|
| Tour guide | Leading visitors through facilities and exhibitions |
| Shopping guide | Assisting retail customers with product information |
| Reception | Greeting and directing guests at offices and hotels |
| Sales assistant | Supporting product demonstrations in XPENG showrooms |
| Security patrol | Monitoring facilities and performing inspection rounds |
The company intentionally defers residential deployment due to safety concerns and the technical challenges of navigating cluttered home environments.[5]
XPENG Robotics has announced a strategic partnership with Baoshan Iron & Steel (Baosteel), one of China's largest steel producers. Baosteel chairman Zou Jixin confirmed that IRON robots will be deployed at Baosteel facilities to "explore application scenarios and iterate and evolve in complex industrial fields such as inspection." The partnership will focus on equipment inspection and monitoring, detecting wear and faults before breakdowns occur. This represents a selective industrial application where mobility and visual inspection capabilities are more important than manipulation.[16][30]
To foster broader adoption and application development, XPENG Robotics has announced plans to release IRON's software development kit (SDK) to global developers. This open-ecosystem approach mirrors successful models in mobile and automotive software and is intended to position IRON as a platform rather than a standalone product, accelerating the creation of third-party applications and services.[16]
On January 31, 2026, IRON made its first public appearance at Shenzhen Bay MixC shopping center. After performing a smooth catwalk to the center of the audience, the robot lost its balance while standing still and fell backward. The incident went viral on Chinese social media. He Xiaopeng responded on Weibo: "It reminds me of how all toddlers learn to walk. After a fall, they will stand firm; the next step is to begin running, and to keep running." The following day, IRON returned to the event using a support frame while engaging audiences in educational exchanges and trivia.[31][32]
XPeng positions itself as a "Physical AI" company rather than a traditional automaker, integrating advanced artificial intelligence into physical machines across three domains:
| Domain | Product | Status |
|---|---|---|
| Smart EVs | P7+, G9, G6, X9, MONA M03 | Mass production |
| Humanoid robots | XPENG IRON | Mass production targeted end of 2026 |
| Flying cars | Land Aircraft Carrier (AeroHT) | Mass production commenced late 2025; 6,000+ pre-orders |
| Robotaxi | Three self-developed models | Trial operations in 2026 |
He Xiaopeng announced at the 2025 AI Day that XPeng's strategic positioning had been upgraded to "a mobility explorer in the physical AI world and a global embodied intelligence company." This framing places robotics at the center of XPeng's long-term identity, alongside rather than subordinate to its EV business.[16]
The convergence strategy is underpinned by shared technology infrastructure. XPeng designs its own AI chips (Turing), develops vision-language-action models (VLA 2.0), manufactures electric motors and batteries at scale, and operates a cloud computing cluster with 2.51 EFLOPS of computing power for AI model training. All of these assets serve both the automotive and robotics divisions simultaneously.[25][26]
XPENG Robotics operates in a rapidly growing market for humanoid robots, particularly in China, where the Chinese government has announced plans to build a world-class humanoid robot industry by 2027. Several other Chinese EV manufacturers and technology companies have entered the space.
| Company | Robot | Country | Notable features |
|---|---|---|---|
| Tesla | Optimus | United States | Iterative development; factory deployment; global brand |
| Unitree Robotics | G1, H1, H2 | China | Low entry price (~$16,000 for G1); strong agility demonstrations |
| UBTECH Robotics | Walker S | China | Early mover in humanoid service robots |
| Agibot | A2 | China | Backed by Shanghai AI Lab; modular design |
| EngineAI | SE01 | China | Founded by former XPENG Robotics team member Zhao Tongyang |
| Figure AI | Figure 02 | United States | BMW factory deployment; OpenAI partnership |
| Xiaomi | CyberOne | China | Consumer electronics ecosystem integration |
XPENG Robotics differentiates itself on three pillars: its highly anthropomorphic design philosophy (prioritizing human likeness over industrial efficiency), the all-solid-state battery (a safety and energy density advantage), and the deep integration with an established automotive manufacturing and AI ecosystem. Where Tesla's Optimus emphasizes functionality and iterative hardware development, XPENG has pursued biomimicry so precise that audiences initially could not distinguish IRON from a human performer.[16][33]
The connection to EngineAI is particularly notable: EngineAI was founded by Zhao Tongyang, the original Dogotix founder who left XPENG Robotics after internal disagreements. Both companies trace their technical origins to the same quadruped robotics team established in Shenzhen in 2016.[6]