Xiaomi CyberDog
Last reviewed
May 2, 2026
Sources
15 citations
Review status
Source-backed
Revision
v2 · 4,002 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 2, 2026
Sources
15 citations
Review status
Source-backed
Revision
v2 · 4,002 words
Add missing citations, update stale details, or suggest a clearer explanation.
Xiaomi CyberDog (Chinese: 铁蛋, pinyin Tiědàn, literally "Iron Egg") is a family of bio-inspired quadruped robots developed by Xiaomi's Robotics Lab in Beijing. The first generation, CyberDog, was announced on August 10, 2021. A second generation, CyberDog 2, followed on August 14, 2023. Both robots are positioned as low-cost, developer-oriented platforms running on NVIDIA Jetson Xavier NX hardware with a Robot Operating System software stack released under the Apache 2.0 license on GitHub.
The CyberDog line sits in the same product category as Boston Dynamics' Spot, Unitree's Go1 and Go2, and ANYbotics' ANYmal. It targets a different audience: while Spot is a $74,500 industrial inspection platform, CyberDog launched at 9,999 RMB (about $1,540 USD) and CyberDog 2 at 12,999 RMB (about $1,785 USD), placing both within reach of university labs, hobbyist roboticists, and individual developers. Xiaomi has not pitched the CyberDog as a finished consumer product, and the first generation was sold as an explicit "engineer edition" limited to 1,000 units.
The project is positioned by Xiaomi as evidence that consumer-electronics scale and supply-chain efficiency can collapse the price gap between research-grade legged robots and the rest of the consumer hardware market. Xiaomi CEO Lei Jun has used CyberDog as a recurring example in his annual public speeches when arguing that the company's manufacturing depth gives it an edge in the early robotics market.
| Manufacturer | Xiaomi (Xiaomi Robotics Lab, Beijing) |
| Chinese name | 铁蛋 (Tiědàn, "Iron Egg") |
| First announced | August 10, 2021 (CyberDog 1) |
| Second generation | August 14, 2023 (CyberDog 2) |
| Launch price (gen 1) | 9,999 RMB (about $1,540 USD), 1,000-unit engineer edition |
| Launch price (gen 2) | 12,999 RMB (about $1,785 USD) |
| Compute | NVIDIA Jetson Xavier NX (8 GB, 21 TOPS) |
| Operating system | Ubuntu 18.04 with ROS 2 Galactic |
| Open source | github.com/MiRoboticsLab (Apache 2.0) |
| Related product | CyberOne humanoid robot (announced August 11, 2022) |
CyberDog was unveiled by Xiaomi at a Beijing launch event on August 10, 2021, the same evening as the Mi MIX 4 smartphone. The company described it as a "bio-inspired quadruped robot" intended for the open source community and developers. The first 1,000 units were sold to Xiaomi fans, engineers, and robotics enthusiasts at 9,999 RMB. Xiaomi explicitly said it was not selling the unit at a profit and pitched the device as a way to seed an external developer community around its in-house robotics work.
The Chinese product name, 铁蛋 (Tiědàn, "Iron Egg"), is a folk nickname commonly given to boys in northern Chinese rural families. Lei Jun and Xiaomi staff have used the nickname interchangeably with "CyberDog" in marketing and on Lei Jun's Weibo account, which gives the robot a more domestic, less industrial framing than its Western product name suggests.
The body uses a magnesium alloy chassis and weighs roughly 14 kg, with overall dimensions of about 560 mm long, 340 mm wide, and 480 mm tall when standing. Each of the four legs has three degrees of freedom, driven by Xiaomi's in-house servo motors with a peak torque of 32 N·m and a peak rotation of 220 RPM. The robot can reach a top speed of 3.2 m/s (about 11.5 km/h) and is capable of dynamic moves including backflips and stair climbing. Maximum payload is 3 kg.
A single 6S lithium-polymer battery sits in the center of the chassis. The robot is rated for a runtime of about an hour under normal trotting loads. Sensor and compute hardware sit on top of the chassis, with a small AI camera mounted at the head and a depth and fisheye stack pointing forward. The battery is user-swappable, which makes the platform easier to use in long demonstration sessions where charging downtime would otherwise be limiting.
Xiaomi described the perception suite as "11 high-precision sensors," which includes touch sensors, ultrasonic sensors, GPS, an Intel RealSense D450 depth camera, binocular ultra-wide-angle fisheye cameras, an AI interactive camera, and a six-microphone array for voice input. The combination supports SLAM, autonomous navigation, obstacle avoidance, person following, gesture recognition, and voice command via Xiaomi's XiaoAi assistant (Chinese only at launch).
The compute backbone is an NVIDIA Jetson Xavier NX module: a six-core 64-bit Carmel ARM CPU, a 384-core Volta GPU with 48 Tensor Cores, two NVDLA deep learning accelerators, 8 GB of LPDDR4 RAM, and 21 TOPS of AI performance at a 15 W TDP. Storage is provided by a 128 GB near-industrial-grade SSD. External I/O includes three USB Type-C ports and one HDMI port, intended for plugging in additional sensors such as LiDAR or extra cameras.
Most coverage focused on price. The Robot Report noted that Xiaomi was charging "about 1/10th the cost of the Boston Dynamics Spot robot." TechCrunch called it Xiaomi's "creepy robot dog," noting that the design "clearly just leaned in and went full-on Robocop" rather than emulating a friendlier consumer form factor like Sony's Aibo. The robot was widely compared to Spot, with several outlets pointing out that the limited 1,000-unit allocation and Chinese-only voice control were major caveats for international buyers.
CyberDog 2 was announced on August 14, 2023, the day before Xiaomi CEO Lei Jun's annual speech, alongside the Xiaomi MIX Fold 3 smartphone. It was priced at 12,999 RMB (around $1,785 USD) and was no longer billed as a limited-run engineer edition, though purchases remained largely China-restricted. Lei Jun used a chunk of his 2023 annual speech to walk through the development history of the CyberDog line, framing the second generation as proof that Xiaomi's robotics group had matured into a self-sustaining engineering effort.
The most obvious change is the look. Where the original CyberDog was a metallic exoskeleton recalling Spot, CyberDog 2 is styled after a Doberman pinscher: matte black body, a more compact silhouette, and a head that swivels to follow people. It is smaller and lighter than its predecessor at 562 mm long, 339 mm wide, and 481 mm standing tall, and weighs about 8.9 kg. Maximum payload is 1 kg. The new shell is built from a custom plastic composite rather than the magnesium alloy used on the original, which contributed to the weight reduction.
The motors were redesigned. The new "CyberGear" micro-actuators are about 40% smaller in volume than the first-generation servos and have, by Xiaomi's account, single-motor torque accuracy 50% higher than the previous generation, with a peak torque density of 37.85 N·m/kg. Each leg retains three degrees of freedom for a total of 12 DOF across the body. Xiaomi advertises a maximum walking speed of 1.6 m/s, with running speeds quoted up to 3.2 m/s in some demonstrations. The robot can backflip continuously, recover from a fall on its own, and step up onto a moving skateboard.
Sensor count rose from 11 to 19. The added hardware is concentrated around what Xiaomi calls a fusion sensing and decision-making system. The robot carries an RGB camera, an AI interactive camera, a binocular fisheye, an Intel RealSense D430 depth camera, a single-line LiDAR, four time-of-flight (ToF) sensors arranged for short-range obstacle detection, an ultrasonic sensor, contact and force sensors on the limbs, a touch surface on the back of the head, and a four-microphone array for far-field voice. Together these feed Xiaomi's bionic motion control algorithms and what the company markets as an AI Computing Vision Engine plus a voice semantic engine.
The compute platform is unchanged in family. CyberDog 2 still runs on an NVIDIA Jetson Xavier NX (8 GB RAM, 21 TOPS at 15 W) supplemented by two co-processors. Storage is listed at 48 GB. External I/O is similar to the first generation: three USB Type-C ports and one HDMI port for developer expansion. Operating system is Ubuntu 18.04 with ROS 2 Galactic.
The February 2024 TechCrunch coverage of CyberDog 2's wider international roll-out noted that the robot was being demonstrated abroad at events such as MWC 2024 in Barcelona, but that its primary buyers remained Chinese universities and research labs. CyberDog 2 also added a set of pre-canned "emotion" expressions through programmable LEDs and head movement: tail wagging, head tilts, and facial-style LED patterns intended to make the robot more legible as a companion device rather than a research tool.
| Specification | CyberDog (2021) | CyberDog 2 (2023) |
|---|---|---|
| Launch date | August 10, 2021 | August 14, 2023 |
| Launch price | 9,999 RMB (about $1,540 USD) | 12,999 RMB (about $1,785 USD) |
| Initial run | 1,000 units (engineer edition) | Open sale in China |
| Weight | About 14 kg | About 8.9 kg |
| Length | 560 mm | 562 mm |
| Width | 340 mm | 339 mm |
| Standing height | 480 mm | 481 mm |
| Payload | 3 kg | 1 kg |
| Max walking speed | 3.2 m/s | 1.6 m/s walking, up to 3.2 m/s running |
| Vertical leap | 0.3 m | 0.2 m |
| Max climb angle | 20° | 20° |
| Degrees of freedom | 12 (3 per leg) | 12 (3 per leg) |
| Servo motor peak torque | 32 N·m | Not officially stated; 37.85 N·m/kg torque density |
| Compute | NVIDIA Jetson Xavier NX, 8 GB RAM, 21 TOPS | NVIDIA Jetson Xavier NX + 2 co-processors, 8 GB RAM, 21 TOPS |
| Storage | 128 GB SSD | 48 GB |
| Battery runtime | About 60 min | About 90 min |
| Sensors | 11 | 19 |
| External I/O | 3x USB-C, 1x HDMI | 3x USB-C, 1x HDMI |
| Operating system | Ubuntu 18.04 / ROS 2 Galactic | Ubuntu 18.04 / ROS 2 Galactic |
| Sensor | CyberDog | CyberDog 2 |
|---|---|---|
| AI interactive camera | Yes | Yes |
| Binocular ultra-wide fisheye cameras | Yes | Yes |
| RGB camera | Bundled with AI camera | Yes (separate) |
| Intel RealSense depth module | D450 | D430 |
| LiDAR | Optional add-on | Single-line LiDAR (built in) |
| Time-of-flight (ToF) sensors | One | Four |
| Ultrasonic sensor | Yes | Yes |
| Touch sensor | Yes (head) | Yes (back of head) |
| Force sensors | No | Yes (limbs) |
| GPS module | Yes | Yes |
| Microphone array | 6-mic | 4-mic |
| IMU | Yes | Yes |
| Total sensors | 11 | 19 |
Both CyberDogs run Ubuntu 18.04 on the NVIDIA Jetson Xavier NX. The high-level software is built on ROS 2 Galactic with Cyclone DDS as the middleware. NVIDIA's Linux for Tegra (L4T) BSP for Xavier NX shipped only with Ubuntu 18.04 at the time, which forced the team to backport ROS 2 Galactic to that release. The repository documents a custom ROS 2 fork that drops dependencies the team did not need.
The CyberDog stack is divided into a few main process groups: motion control, perception, decision, and connectivity. Motion control runs at the lowest level on a real-time loop and talks to the leg motors through an EtherCAT-style bus. Perception nodes consume the depth, fisheye, RGB, and AI camera streams, run obstacle detection, SLAM, and person tracking, and publish results onto ROS 2 topics. The decision layer translates user goals (follow-me, navigation waypoints, behavior triggers like "sit") into motion plans. A connectivity node handles the smartphone app, voice input, and cloud services.
Xiaomi's voice stack on the original CyberDog is XiaoAi, the assistant that ships in Xiaomi smartphones and smart speakers, which is closed source. The repository ships a stub for the voice node and a build flag to swap in the closed-source binary for users with Xiaomi accounts. Deep learning workloads on the device include a person detector, a face recognizer for owner following, gesture recognition, an audio direction-of-arrival estimator for the microphone array, and a depth-completion network on the RealSense streams. None of these models is unusually large; the Xavier NX has a hard 8 GB memory limit and the team was clearly conscious of the platform's TOPS budget.
For CyberDog 2, the same general architecture survives, but the perception graph grew to handle the new sensors (LiDAR, four ToF, force sensors), and Xiaomi added bionic motion control algorithms that draw on reinforcement-learning-trained gait policies. The public repository has more locomotion code than perception code.
The locomotion controller on CyberDog 2 is the first generation of the platform that publicly leans on reinforcement learning. Xiaomi engineers have described training gait policies in NVIDIA Isaac Gym and then transferring them to the physical robot through a sim-to-real pipeline that mixes domain randomization with on-robot calibration. The pipeline borrows ideas from the academic work that produced ETH Zurich's ANYmal RL controllers and MIT's Mini Cheetah locomotion stack. Public technical talks by the Xiaomi Robotics Lab in 2023 and 2024 walked through how the company structured the reward function around stability, energy efficiency, and recovery from disturbances.
Reinforcement-learning-based gait control enables some of the more demonstrative behaviors in CyberDog 2's marketing reels. The robot can recover from being shoved off a table, walk while balancing on a moving skateboard, and chain backflips without manually scripted choreography. CyberDog 2 was one of the first sub-$2,000 quadrupeds to ship those capabilities as off-the-shelf behaviors rather than research demos.
The core software for CyberDog is published at github.com/MiRoboticsLab. The main repository, cyberdog_ros2, contains the ROS 2 packages for the first generation, including motion control, navigation, multi-modal perception, and target tracking. It is licensed under Apache 2.0. Companion repositories cover the simulator (cyberdog_simulator) and the visual-inertial SLAM stack (cyberdog_mivins). For CyberDog 2, Xiaomi has published a different set of repositories with a redesigned application framework, also under Apache 2.0.
In practice the open source release is partial. Motion control, the gait library, and the ROS 2 plumbing are open. The voice assistant binary, parts of the cloud app integration, and some of the safety-critical motor firmware are closed. This is similar to what Unitree ships with the Go1: enough open code to replace the high-level stack, but not a fully open hardware platform.
A small community has formed around the platform. Notable third-party work includes the MAVProxyUser/ConsumerQuadruped repository, which collects reverse-engineering notes for running custom code on the device, and several university projects that use CyberDog as a teaching platform for legged locomotion classes. The community is smaller than Unitree's Go1 community, partly because CyberDog 1 was capped at 1,000 units and international buyers have had difficulty acquiring either generation. Xiaomi has run "CyberDog Open Source" community competitions and developer events in mainland China, including a 2022 hackathon co-hosted with Tsinghua University.
| Robot | Maker | Launch | Approx. price | Weight | Top speed | Compute | Open source |
|---|---|---|---|---|---|---|---|
| CyberDog | Xiaomi | 2021 | $1,540 USD | 14 kg | 3.2 m/s | NVIDIA Jetson Xavier NX | ROS 2 stack (Apache 2.0) |
| CyberDog 2 | Xiaomi | 2023 | $1,785 USD | 8.9 kg | 1.6 m/s walking | NVIDIA Jetson Xavier NX | ROS 2 stack (Apache 2.0) |
| Spot | Boston Dynamics | 2019 (commercial) | $74,500 USD | 33 kg | 1.6 m/s | Atom CPU + custom controllers | Closed (SDK available) |
| Unitree Go1 | Unitree | 2021 | from $2,700 USD | 12 kg | 4.7 m/s | Custom + optional NVIDIA Nano | Partial SDK |
| Unitree Go2 | Unitree | 2023 | from $1,600 USD | 15 kg | 3.7 m/s | NVIDIA Jetson Orin (EDU model) | Partial SDK |
| ANYmal C | ANYbotics | 2019 | Six figures USD (industrial) | 50 kg | 1.0 m/s | Industrial PC | Closed |
| MIT Mini Cheetah | MIT Biomimetics Robotics Lab | 2019 (research) | Not for sale | 9 kg | 2.45 m/s | NVIDIA Jetson TX2 | Open hardware design |
Within this group, CyberDog occupies a niche between MIT's Mini Cheetah research platform and Unitree's commercial Go-series. It is more polished than a research robot and more available than Spot or ANYmal, but the limited release and China-restricted distribution have kept it from displacing Unitree as the default low-cost quadruped for academic work.
Launch coverage was a mix of admiration for the price-to-capability ratio and skepticism about how serious the product was. New Atlas headlined its piece "Xiaomi joins the pack with $1,500 CyberDog robot that does backflips," noting the robot was "about 1/10th the cost of Boston Dynamics' Spot." The Robot Report described it as a developer platform aimed at hobbyists who would otherwise be priced out of legged robotics. CNX Software focused on the Jetson Xavier NX integration and called the price "impressive."
More critical takes came from outlets focused on consumer hardware. TechCrunch's August 2021 piece called the dog "creepy" and questioned whether the form factor was something most people would want. Android Police wrote that "the Xiaomi CyberDog is an awesome vision of our creepy smart home future." Several reviewers pointed out that voice control was Mandarin-only at launch and the smartphone app was tied to a Xiaomi account. IEEE Spectrum's robot directory added entries for both CyberDog and CyberDog 2, noting Xiaomi as one of the few major Chinese consumer-electronics firms to ship a real legged robot for sale.
For CyberDog 2, the reception shifted. The redesigned Doberman aesthetic was received better than the metallic look of the first generation, and reviewers praised the new motors and smoother gait. Designboom highlighted demonstrations of the robot performing ballet steps and a moonwalk; Interesting Engineering focused on the 19-sensor perception stack. The price increase from 9,999 RMB to 12,999 RMB drew some grumbling, but most reviewers concluded it was still a strong value. The Verge and Engadget both ran short coverage pieces during MWC 2024, where Xiaomi brought a small unit to its booth.
The CyberDog line is one product of the Xiaomi Robotics Lab, which Xiaomi established explicitly to develop in-house robotics IP. The same group went on to launch CyberOne, a 1.77 m, 52 kg humanoid robot, in August 2022, and has continued to publish progress reports on bionic motion control and humanoid manipulation since.
Xiaomi's broader pitch is that robotics is a natural extension of its consumer electronics business. The company already builds at scale, has a tight supply chain, and ships software via its MIUI/HyperOS ecosystem. Lei Jun has argued in his annual speeches that bringing those capabilities to legged robots and humanoids should let Xiaomi reach prices an order of magnitude below incumbent industrial vendors. CyberDog and CyberDog 2 are the public proof points for that thesis.
The CyberDog and CyberOne projects share several engineering reuse points. The CyberGear actuators that debuted on CyberDog 2 are also used in CyberOne's joints, with different gearing ratios and torque envelopes. The same motion-control software stack drives both lines, with the bipedal balance code on CyberOne reusing controllers that were originally written and tuned on the quadruped. Xiaomi has been clear in technical papers and interviews that the CyberDog program is a deliberate stepping stone toward humanoid manipulation, and that lessons learned on quadrupeds, particularly around low-cost actuator manufacturing, are intended to flow upward into the CyberOne line.
Whether the strategy translates into a mass-market consumer product is unclear. Both CyberDogs are still developer hardware; neither has the kind of polished out-of-the-box experience that a typical Xiaomi customer expects from a phone or vacuum. The third generation, if it appears, will be the more interesting test of whether Xiaomi can make a quadruped that ships beyond the engineer edition.
Documented uses of the CyberDog platform fall into a few buckets:
CyberDog and CyberDog 2 share several limitations that have kept them from broader adoption. The platforms have limited general-purpose autonomy: the follow-me, voice, and obstacle-avoidance behaviors are scripted on top of canned policies, and the robot does not generalize to arbitrary tasks the way an LLM-driven mobile robot might. Voice and app integration are tied to Xiaomi's ecosystem and Mandarin-language services, which makes the platform awkward for non-Chinese-speaking users. Distribution is concentrated in mainland China, with international buyers largely relying on resellers or grey-market imports because there is no official global storefront. The motor and battery are not designed for high-load industrial work; maximum payload of 1 to 3 kg restricts the platform to inspection-style applications and rules out heavier tool-carrying tasks that Spot can do. Finally, the sim-to-real and reinforcement learning training pipeline used internally at Xiaomi has not been open sourced, so researchers can use the runtime stack but have to bring their own training infrastructure to develop new gait policies.
Xiaomi has not publicly announced a CyberDog 3 as of early 2026. The CyberDog 2 platform has continued to receive software updates through Xiaomi's developer portal, including incremental improvements to the gait policy and additional pre-canned interaction behaviors. Xiaomi Robotics Lab has signaled in interviews and conference talks that future investment is concentrated on the CyberOne humanoid line rather than another CyberDog generation, on the argument that legged manipulation is the more open research problem.
Media coverage in 2024 noted that CyberDog 2 saw expanded international demonstration appearances at MWC 2024 and at robotics events in Japan and Europe, but Xiaomi has not opened formal export sales channels.