CtrlK
The Galaxea R1 Lite is a compact wheeled dual-arm [[humanoid robot]] developed by [[Galaxea Dynamics|Galaxea AI]] (also known as Galaxea Dynamics), a Chinese [[embodied AI]] and [[robotics]] startup founded in September 2023. Positioned as the entry-level and most compact model in the company's [[Galaxea R1|R1 series]], the R1 Lite is designed primarily as a data collection platform for training embodied AI models, as well as a general-purpose research and education robot. It stands 128 cm tall, weighs 55 kg, and features 23 [[degrees of freedom]] distributed across a three-wheel omnidirectional chassis, a 3-DOF torso, and two 6-DOF Galaxea A1X robot arms equipped with G1 parallel grippers.
The R1 Lite was announced in January 2025 alongside the [[Galaxea R1 Pro|R1 Pro]] as part of an expansion of the R1 product lineup. While the standard R1 and R1 Pro target industrial manipulation and complex service tasks, the R1 Lite occupies a distinct niche: it provides a lower-cost, lighter-weight platform optimized for teleoperated data collection, algorithm development, and laboratory-scale research. The R1 Lite has been used as the primary data collection embodiment for the Galaxea Open-World Dataset, a 500-hour corpus of real-world manipulation demonstrations that underpins the company's G0 family of Vision-Language-Action (VLA) models.
[[Galaxea Dynamics|Galaxea AI]] was founded in September 2023 by a team of scientists with backgrounds at [[Tsinghua University]], Stanford, MIT, and UC Berkeley. The company is headquartered in Beijing, China. CEO and founder Gao Jiyang (born 1992) studied electronic engineering at Tsinghua, completed a Ph.D. in computer vision at the University of Southern California, and previously worked at [[Waymo]] and Momenta before starting Galaxea. The company's co-founders include Xu Huazhe, a tenure-track assistant professor at Tsinghua's Institute for Interdisciplinary Information Sciences (IIIS) who leads the Tsinghua Embodied AI Laboratory (TEA Lab); Hang Zhao, a Ph.D. graduate of MIT; and Tianwei Li, formerly a senior director at Momenta.[1][9]
Galaxea AI's stated mission is to build "embodied intelligence at a global scale: 10 billion robots for 10 billion people." By April 2026, the company had raised over $500 million in cumulative funding, with its Series B+ round of approximately $291 million pushing its valuation above $2.9 billion.[7][8] The company serves more than 40 clients worldwide, including [[Huawei]] Cloud, [[Volkswagen]], Haier, [[Samsung]], [[ByteDance]], Physical Intelligence, [[Stanford University]], and MIT.[5]
The R1 series consists of three wheeled dual-arm humanoid robots that share a common design philosophy: a human-like upper body with dual manipulator arms mounted on a wheeled mobile base. This approach prioritizes practical deployability in structured indoor environments over the bipedal locomotion used by robots like [[Tesla Optimus]] or [[Boston Dynamics]]' [[Atlas (robot)|Atlas]].
The three models in the series are:
| Model | Primary role | Total DOF | Arm configuration | Max height | Approx. price |
|---|---|---|---|---|---|
| [[Galaxea R1 | R1]] (Standard) | General-purpose mobile manipulation | 24 | Dual 6-DOF (A1 arms) | 200 cm |
| [[Galaxea R1 Pro | R1 Pro]] | Industrial collaboration and complex tasks | 26 | Dual 7-DOF (A2 arms) | 200 cm |
| R1 Lite | Data collection and research | 23 | Dual 6-DOF (A1X arms) | 170 cm | ~$40,000 USD |
The R1 Lite was introduced to address a specific gap in the market: researchers and AI laboratories needed an affordable, compact mobile manipulation platform for collecting real-world training data at scale, without the full sensor suite and industrial payload capacity of the standard R1 or R1 Pro.[2][3]
The R1 Lite has a notably more compact form factor than the other R1 variants. It stands 1,281 mm (128.1 cm) tall in its default standing posture, compared to the 1,695 mm (169.5 cm) height of the standard R1 and R1 Pro. Its chassis measures 596 mm long by 600 mm wide by 250 mm high, while the upper platform is 670 mm long by 298 mm wide. The total weight, including the battery, is 55 kg, making it roughly 41 kg lighter than the 96 kg standard R1.[4][13]
This smaller footprint is intentional. At 670 mm overall width, the R1 Lite can navigate through narrow doorways, crowded laboratory spaces, and residential interiors more easily than its larger siblings. The reduced height also means the robot's workspace is better matched to tabletop-height manipulation tasks, which constitute a large fraction of the tasks in typical research settings.
The R1 Lite has 23 total degrees of freedom, distributed as follows:[4]
| Subsystem | DOF | Details |
|---|---|---|
| Chassis | 6 | Three self-developed steering wheel modules (W1), each contributing 2 DOF |
| Torso | 3 | Waist yaw (plus or minus 170 degrees), hip pitch (plus or minus 100 degrees), and vertical lift |
| Left arm with gripper | 7 | 6-DOF A1X arm + 1-DOF G1 gripper |
| Right arm with gripper | 7 | 6-DOF A1X arm + 1-DOF G1 gripper |
| Total | 23 |
The 23-DOF count is one fewer than the standard R1 (24 DOF) and three fewer than the R1 Pro (26 DOF). The difference from the standard R1 comes from the torso: the R1 Lite has a 3-DOF torso while the standard R1 has a 4-DOF torso. The difference from the R1 Pro is further compounded by the Pro's use of 7-DOF A2 arms rather than 6-DOF arms.
The R1 Lite's omnidirectional chassis uses three self-developed W1 steering wheel modules, providing 6 degrees of freedom of motion. This vector-drive design supports multiple locomotion modes: translation in any direction, spinning in place, and Ackermann-style steering for smooth curved trajectories. The maximum linear speed is 1.5 m/s (5.4 km/h).[4][13]
The 360-degree unlimited rotation capability of each steering module allows the robot to move sideways, rotate on the spot, or follow complex curved paths without needing to reorient its body. This is particularly useful in cluttered laboratory and household environments where the robot may need to approach objects from various angles.
The R1 Lite is equipped with two Galaxea A1X robot arms, each with 6 degrees of freedom. The A1X arms differ from the A1 arms used in the standard R1; they are optimized for the Lite platform's smaller form factor while maintaining comparable manipulation capabilities. Each arm measures 600 mm in length when fully extended and 385 mm when folded, with a width of 100 mm. The individual arm weight is 4.2 kg.[4]
The six joints of each arm are equipped with high-precision, high-torque planetary gear motors that allow independent variable-speed operation. The arms feature spherical wrist joints that mimic human wrist flexibility, enabling the robot to approach and grasp objects from a wide range of angles.
| Arm specification | Value |
|---|---|
| Arm model | Galaxea A1X |
| DOF per arm | 6 |
| Extended length | 600 mm |
| Folded length | 385 mm |
| Arm weight | 4.2 kg |
| Rated payload (at 0.6 m) | 3 kg |
| Maximum payload (at 0.6 m) | 5 kg |
| Working height range | 0 to 1.7 m |
Each arm is fitted with a Galaxea G1 parallel gripper. The G1 provides a rated gripping force of 100 N and a stroke (opening width) of 0 to 100 mm, suitable for grasping a wide variety of household and laboratory objects.[4]
The 3-DOF torso provides waist yaw rotation of plus or minus 170 degrees, hip pitch of plus or minus 100 degrees, and vertical lift that extends the robot's working height from ground level up to 1.7 meters. The torso motors provide a rated torque of 108 Nm and a maximum torque of 304 Nm. The two knee joints have motion ranges of 0 to 100 degrees (W1) and negative 154 to 145 degrees (W2).[4]
The vertical lift capability is a distinctive feature of the R1 series design. Rather than relying solely on arm reach, the robot can physically raise or lower its entire upper body to access objects at different heights, from floor-level items to objects on high shelves.
One of the most notable differences between the R1 Lite and the other R1 models is the computing platform. While the standard R1 and R1 Pro use [[NVIDIA]] Jetson AGX Orin 32GB systems with 200 TOPS of GPU compute, the R1 Lite uses an Intel Core i9-12900HK processor. This 14-core CPU (with a base clock of 2.5 GHz and boost up to 5.0 GHz) provides strong general-purpose computing performance but lacks the dedicated GPU accelerator present in the Jetson-based variants.[4]
| Computing specification | Value |
|---|---|
| Processor | Intel Core i9-12900HK (14-core, 2.5 GHz base, 5.0 GHz boost) |
| Memory | 32 GB LPDDR4 |
| Storage | 1 TB SSD |
| Power consumption | 19 V, 6.32 A |
| NUC weight | 358 g |
| USB ports | 2x USB 2.0, 2x USB 3.0, 1x Type-C |
| Display output | 2x HDMI |
| Cooling | Smart fan with copper heat sink |
| Wireless | WiFi 6 (AX 201), Bluetooth 5.2 |
The choice of an Intel NUC-class computing unit rather than an NVIDIA Jetson reflects the R1 Lite's primary role as a data collection and teleoperation platform rather than an edge inference platform. During teleoperated data collection, the computational load is primarily on sensor data recording and communication rather than real-time neural network inference. For applications requiring on-device AI inference, researchers can offload computation to an external workstation connected over the network. The R1 Lite's WiFi 6 and Ethernet connectivity support this workflow.[4]
The R1 Lite runs Ubuntu 22.04 LTS with ROS 2 Humble as its primary middleware layer. The software stack, called "mobiman" (short for mobile manipulation), provides a layered architecture with hardware abstraction, motion control, and high-level application interfaces.[10][11]
Key software modules include:
The software supports both ROS 1 (Noetic) and ROS 2 (Humble) versions, though the platform has transitioned to ROS 2 as the primary supported distribution. Galaxea provides an SDK (the Galaxea ATC ROS2 SDK) with regular updates; version 2.1.4 was the latest release as of early 2026.[10][11]
Galaxea AI maintains an active open-source presence through its OpenGalaxea GitHub organization. Resources relevant to the R1 Lite include:
These resources lower the barrier for laboratories and startups to adopt the R1 Lite as a research platform, since researchers can train and test control policies in simulation before deploying them on the physical robot.
The R1 Lite has a more streamlined sensor suite compared to the standard R1 and R1 Pro, in keeping with its focus on cost-effective data collection.
| Camera | Type | Resolution | Frame rate | Field of view | Depth range | Weight |
|---|---|---|---|---|---|---|
| Platform (head) | Binocular stereo | 1280 x 720 per eye (2560 x 720 combined) | 15/30/60 FPS | 126 degrees H x 116 degrees V | 0.25 to 10 m | 50 g |
| Left wrist | Monocular depth (Intel RealSense D405) | 1280 x 720 | 30 FPS | 87 degrees H x 58 degrees V x 95 degrees D | 0.2 to 3 m | 75 g |
| Right wrist | Monocular depth (Intel RealSense D405) | 1280 x 720 | 30 FPS | 87 degrees H x 58 degrees V x 95 degrees D | 0.2 to 3 m | 75 g |
The platform camera, mounted on the robot's head, provides a stereo RGB image for scene-level context and depth estimation. The two Intel RealSense D405 wrist cameras provide close-range depth perception during manipulation tasks, with an effective depth range of 0.2 to 3 meters. This dual-camera configuration on the wrists is particularly useful for precise grasping and object manipulation, as the wrist-mounted viewpoint moves with the end effector.[4][6]
Notably, the R1 Lite does not include the five chassis-mounted monocular cameras found on the standard R1 and R1 Pro. It also lacks the dedicated [[LiDAR]] sensor present on the other models, although some third-party listings reference optional LiDAR integration. The reduced camera count contributes to the lower cost and simpler system architecture.
| Sensor | R1 Lite | R1 (Standard) | R1 Pro |
|---|---|---|---|
| Head camera | 1 binocular stereo (720p) | 1 binocular stereo (1080p) | 2 monocular (1080p) |
| Wrist cameras | 2 monocular depth | 2 monocular depth | 2 monocular depth (optional) |
| Chassis cameras | None | 5 monocular (1080p) | 5 monocular (1080p) |
| LiDAR | None (optional) | 1x 360-degree | 1 to 2x 360-degree |
| Force sensors | No | No | Integrated |
| IMU | N/A | Yes | Yes |
The R1 Lite uses a 48 V lithium-ion battery pack with a 15 Ah capacity, providing 720 Wh of total energy storage. The battery includes a Battery Management System (BMS) for safe charging and discharge management.[4]
The robot supports two power supply modes: battery-powered for untethered mobile operation, and an optional AC power mode (110 to 220 V) for stationary tasks where continuous operation is needed without battery limitations. The dual power mode capability is useful in laboratory settings where the robot may spend extended periods collecting data at a fixed workstation.
The R1 Lite's 720 Wh battery is significantly smaller than the 1,680 Wh battery used in the standard R1 and R1 Pro. This is consistent with the Lite's lower weight (55 kg vs. 96 kg) and reduced power consumption from its lighter arms and fewer sensors. The battery provides approximately three hours of typical operation.
| Power specification | Value |
|---|---|
| Rated voltage | 48 V |
| Battery type | Lithium-ion |
| Capacity | 15 Ah |
| Energy | 720 Wh |
| BMS | Supported |
| AC power option | 110 to 220 V (optional) |
| Estimated runtime | ~3 hours |
Teleoperation is a core capability of the R1 Lite, reflecting its primary purpose as a data collection platform. The robot supports isomorphic teleoperation, in which a human operator controls the robot using a scaled-down physical replica called the R1-T Isomorphic Remote Operation Platform.
The isomorphic teleoperation system maps the operator's arm and body movements directly to the robot's kinematics through full-joint mapping. Because the control platform replicates the robot's physical structure at a smaller scale, the operator's arms naturally stay within reachable postures, avoiding the inverse kinematics failures that can occur with other teleoperation approaches. The system provides force feedback from the robot back to the operator, giving tactile awareness of the manipulation environment.[2][3]
The teleoperation system achieves millimeter-level positional precision and millisecond-level response times. Default chassis velocities during teleoperation are 0.2 m/s for linear motion and 0.6 rad/s for angular motion. Both wired and wireless connections are supported, with wired connections recommended for multi-machine setups to ensure stable data collection.[3]
The R1 Lite supports isomorphic teleoperation but does not include VR teleoperation capability, which is available on the standard R1 and R1 Pro. For many data collection workflows, isomorphic control is preferred over VR because it provides more intuitive force feedback and does not require the operator to wear a headset for extended periods.
The R1 Lite plays a central role in Galaxea AI's approach to developing embodied intelligence. CEO Gao Jiyang has argued that "at least for the next three years, reinforcement learning with simulators will not lead to the ultimate goal" for upper-limb manipulation tasks, and that real-world data combined with [[imitation learning]] is the more promising path.[9] The R1 Lite is the primary tool for collecting this real-world data.
The Galaxea Open-World Dataset, published in September 2025, was collected using a fleet of R1 Lite robots operating across diverse real-world environments. The dataset comprises over 500 hours of high-fidelity manipulation data covering more than 150 distinct tasks across 50 different scenes. The scene categories include residential spaces, retail environments, catering and food service settings, and office workplaces.[6]
All demonstrations in the dataset were recorded using the R1 Lite's 23-DOF embodiment, ensuring uniformity across the dataset. Each demonstration includes synchronized data from the stereo head camera and dual wrist cameras, along with joint positions, velocities, and gripper states. Precise subtask-level language annotations accompany each trajectory to facilitate both training and evaluation of language-conditioned control models.
The data collected by R1 Lite fleets feeds into the training pipeline for Galaxea AI's G0 family of [[Vision-Language-Action]] models. The G0 model uses a dual-system architecture: a Vision-Language Model (G0-VLM) handles high-level task planning and language understanding, while a Vision-Language-Action model (G0-VLA) handles fine-grained motor execution. The two systems operate asynchronously at different frequencies.[6]
G0 training follows a three-stage curriculum: cross-embodiment pre-training using data from multiple robot platforms, single-embodiment pre-training specific to the R1 hardware, and task-specific post-training for particular applications. The R1 Lite's consistent embodiment across the training fleet is important for the single-embodiment stage, as it ensures that the model learns a coherent mapping between visual observations and motor actions.
Subsequent model releases, including G0Plus (January 2026) and G0Tiny (a 250-million-parameter model optimized for edge deployment), build on the foundation of R1 Lite-collected data.
The R1 Lite's primary market is academic and research institutions. Universities and laboratories use the platform as a testbed for learning-based control, dexterous manipulation, human-robot interaction, and [[computer vision]] research. Its ROS 2 compatibility, open-source VLA models, and comprehensive documentation lower the barrier to entry for research groups. The compact form factor makes it practical for laboratory environments where space is limited.[2][13]
Galaxea AI counts [[Stanford University]] and MIT among its research clients, and the company has released simulation assets (URDF models for Isaac Sim, MuJoCo, and Gazebo) that allow researchers to develop and test algorithms in simulation before deploying them on the physical R1 Lite.
As described above, the R1 Lite's most distinctive application is large-scale teleoperated data collection for training embodied AI models. Galaxea has deployed fleets of R1 Lite robots in real-world environments (homes, offices, retail stores, and restaurants) to collect the manipulation demonstrations that power its VLA models. The robot's relatively low cost and compact size make it practical to deploy in multiples, which is necessary for building datasets of the scale required by modern machine learning approaches.[6]
While the R1 Lite is primarily a research and data collection platform, demonstrations have shown it performing practical household tasks including object grasping, transportation, placement, and light assembly work. The 5 kg maximum arm payload and 100 N gripper force are sufficient for handling most common household objects. The vertical operating range of 0 to 1.7 meters covers tasks from picking items off the floor to reaching objects on standard-height shelves.
The R1 Lite is priced at approximately $40,000 USD for international buyers, according to third-party robotics distributors.[13] In China, the R1 series starts at 199,000 RMB (approximately $28,000 USD), though this starting price refers to the standard R1 rather than the R1 Lite. Galaxea co-founder Xu Huazhe has stated that the R1 series robots are priced between $44,500 and $64,000 USD overall, placing the R1 Lite at the lower end of this range.[5]
The R1 Lite is available for pre-order through Galaxea Dynamics' online store and through authorized distributors internationally, including Robots International in North America.
The following table summarizes the principal differences between the three R1 series models:
| Specification | R1 Lite | R1 (Standard) | R1 Pro |
|---|---|---|---|
| Standing height | 1,281 mm | 1,695 mm | 1,700 mm |
| Weight (with battery) | 55 kg | 96 kg | 96 kg |
| Total DOF | 23 | 24 | 26 |
| Torso DOF | 3 | 4 | 4 |
| Arm model | A1X (6-DOF) | A1 (6-DOF) | A2 (7-DOF) |
| Arm reach | 600 mm | 715 mm (863 mm deployed) | 620 mm |
| Max arm payload | 5 kg | 5 kg | 5 kg |
| Operating height | 0 to 170 cm | 0 to 200 cm | 0 to 200 cm |
| Battery energy | 720 Wh | 1,680 Wh | 1,680 Wh |
| Computing platform | Intel i9-12900HK | NVIDIA Jetson AGX Orin 32GB | NVIDIA Jetson AGX Orin 32GB |
| Head camera | Binocular stereo (720p) | Binocular stereo (1080p) | 2 monocular (1080p) |
| Wrist cameras | 2 (depth) | 2 (depth) | 2 (optional depth) |
| Chassis cameras | 0 | 5 | 5 |
| LiDAR | None | 1x 360-degree | 1 to 2x 360-degree |
| Force sensors | No | No | Yes (integrated) |
| Teleoperation | Isomorphic only | Isomorphic + VR | Isomorphic + VR |
| Approx. price | ~$40,000 USD | ~$28,000 USD (199,000 RMB) | $44,500 to $64,000 USD |
| Primary use case | Data collection, research | General-purpose mobile manipulation | Industrial collaboration |
The R1 Lite occupies a specific niche between the other two models. It is lighter and more compact than the standard R1, making it better suited for deployment in space-constrained research environments and for fleet-based data collection. However, its smaller battery, fewer sensors, and lack of a dedicated GPU limit its suitability for fully autonomous industrial deployment. Researchers who need autonomous navigation with LiDAR and chassis cameras, or who require on-device neural network inference, would typically choose the standard R1 or R1 Pro instead.