Dexmate
Last reviewed
Sources
15 citations
Review status
Source-backed
Revision
v5 · 2,541 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
15 citations
Review status
Source-backed
Revision
v5 · 2,541 words
Add missing citations, update stale details, or suggest a clearer explanation.
Dexmate Inc. is an American robotics company based in Santa Clara, California, that builds AI-powered dexterous manipulation robots, most notably Vega, a foldable wheeled humanoid robot that sells from $89,999.[1][7] Founded in 2024 by PhD researchers from MIT, the University of California, San Diego (UCSD), and Carnegie Mellon University (CMU), Dexmate develops dexterous robotic hands and a mobile humanoid platform with omni-directional mobility, aimed at industrial and research customers.[1][2][9] The company is an NVIDIA Inception startup that has raised roughly $33 million in venture funding, with backers including LG Technology Ventures, the corporate venture arm of South Korea's LG Group.[2][9][14]
Dexmate belongs to a wave of US humanoid robotics ventures, alongside companies such as Figure AI, 1X Technologies, and Apptronik, that emerged from elite robot-learning labs and target manufacturing, logistics, and retail customers. Dexmate distinguishes itself by emphasizing dexterous manipulation, a wheeled rather than bipedal lower body, and a hardware-software co-design philosophy in which the physical platform and the machine learning models are developed together.[3][4]
Dexmate is a Silicon Valley robotics startup that designs and manufactures general-purpose mobile manipulator robots for factories, warehouses, and research labs. Its single product line is the Vega robot, a dual-arm humanoid upper body mounted on an omni-directional wheeled base, sold both as a commercial automation platform and as a standard hardware platform for other robotics AI developers.[1][9] According to The Korea Herald, Dexmate's hardware has been adopted as a research hardware platform by global robotics AI developers.[9] The company combines custom hardware with data-driven learning, training its control policies in simulation and from human demonstration before deploying them on real robots.[3][4]
Dexmate was co-founded in 2024 by Tao Chen, Yuzhe Qin, Wenda Wang, and Chongyang (Max) Wang.[3][5] Tao Chen serves as chief executive officer and holds a PhD in Electrical Engineering and Computer Science from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL), where he was advised by Professor Pulkit Agrawal. His doctoral research focused on robot learning, including quadruped locomotion and dexterous manipulation through reinforcement learning. Chen received the Best Paper Award at the Conference on Robot Learning (CoRL) in 2021 for the paper "A System for General In-Hand Object Re-Orientation."[5]
Yuzhe Qin serves as chief technology officer. He earned his PhD in robotics from UCSD, where he was advised by Professor Xiaolong Wang and Professor Hao Su. His research centered on dexterous manipulation, sim-to-real transfer, and learning from human demonstrations, with publications at top venues including CVPR, ICCV, ECCV, ICLR, ICML, CoRL, RSS, ICRA, and IROS. Qin interned at Google X and at NVIDIA from June 2022 to February 2023, and he received a Qualcomm Innovation Fellowship in 2023.[6]
Chongyang Wang serves as chief operating officer. He is described as an MIT graduate with more than a decade of operational experience. Wenda Wang holds a master's degree from CMU and previously worked as a machine learning engineer at Apple. Both Chen and Qin completed undergraduate studies in mechanical engineering at Shanghai Jiao Tong University before moving to the United States for doctoral training. Professor Xiaolong Wang of UCSD, who supervised Qin during his doctoral studies, is publicly affiliated with the company as an advisor.[3][5][6]
The company was incorporated in 2024 and joined NVIDIA's Inception Program, a partner accelerator that gives early-stage AI and robotics startups discounted access to NVIDIA hardware, simulation tools, and go-to-market support. Dexmate built its first general-purpose mobile manipulator, named Vega, in under six months after founding. The team relied heavily on NVIDIA's robotics stack during this period, using Isaac Sim and Isaac Lab for simulation training and Isaac GR00T as a foundation for its vision-language-action models.[2][4] Describing the deployment workflow, Chen told NVIDIA: "When we have a customer request to automate tasks, we go to their factory, we do the scanning, and we create a digital twin and run simulations in Isaac Sim with synthetic data."[2]
Dexmate publicly introduced Vega on March 7, 2025, with a launch video that featured the robot folding its torso, unfolding to overhead reach, manipulating tools, and rolling across a warehouse floor on its omni-directional base. The launch positioned Vega as a research and industrial platform priced significantly below most full-size bipedal humanoids on the market.[7] The company opened pre-orders for Vega in November 2025, with a $999 deposit and a lead time of roughly three months.[13]
In March 2025, Dexmate exhibited Vega at ProMat 2025, the materials handling and logistics trade show held at McCormick Place in Chicago from March 17 to March 20. The booth showcased Vega performing tasks relevant to warehouse and factory operations, including material transportation, inventory handling, and fulfillment-style picking. Several commercial partners were already running pilot tests with the robot at the time of the show, according to Dexmate's exhibitor profile.[8]
During the same period the company demonstrated Vega at NVIDIA's Santa Clara campus, where the robot drew attention by lifting weights to highlight the payload capacity of its arms. The demo coincided with NVIDIA's broader push to spotlight Inception members building on the Isaac stack, and NVIDIA later featured Dexmate in its blog post celebrating two million developers in its robotics ecosystem.[2][4]
Dexmate has raised approximately $33 million in venture funding across multiple rounds, according to startup-data trackers PitchBook and Tracxn.[11][12] Named investors include LG Technology Ventures, Epsilon Ventures, Mana Ventures, RoboStrategy, and Jinqiu Capital.[12]
On March 10, 2026, the South Korean IT services firm LG CNS announced a strategic investment in Dexmate through LG Technology Ventures. Financial terms were not disclosed.[9][10] Tracxn characterizes the LG round as a pre-series stage investment.[12] The deal is part of LG CNS's "Robot Transformation" (RX) business, an effort to integrate hardware from multiple humanoid robotics vendors under a single software and operations platform.[9][14] LG CNS said it would pair Dexmate hardware with a robot foundation model and an operations and training platform, and that it was running proof-of-concept projects across logistics, retail, and manufacturing sites.[14][15] Lee Jun-ho, head of LG CNS's smart logistics and city business unit, said the goal was to "validate the humanoid robot business model that goes beyond technology testing to real-world deployment and lead the era of physical AI."[15]
Vega is Dexmate's flagship mobile humanoid robot, marketed as a general-purpose dual-arm manipulator on a wheeled base. The system is designed to be plug-and-play, requiring no modifications to a customer's existing environment, and ships ready for Python-based programming, Robot Operating System integration, and teleoperation via VR headsets such as the Vision Pro and Meta Quest.[1][7]
| Specification | Details |
|---|---|
| Height (standing) | 171 cm (5 ft 7 in) |
| Height (folded) | 66 cm (2 ft 2 in) |
| Weight | 135 kg |
| Maximum reach | 2.2 m (with extended torso) |
| Walking speed | 4 km/h |
| Mobility | Omni-directional wheeled base |
| Arm DOF | 7 per arm |
| Hand DOF | 12 per hand, 5 fingers (6 DOF option available) |
| Head DOF | 3 |
| Total DOF | 36+ |
| Arm payload | About 7 kg (15 lb) per arm |
| Battery life | 10+ hours under load, up to 20-30 hours light load |
| Sensors | RGB-D cameras, LiDAR, IMUs, ultrasonic, force/torque |
| Compute | Intel x86 CPU + NVIDIA Jetson AGX Orin (32 GB or 64 GB) |
| Operating system | Linux with ROS-compatible SDK |
| I/O | Ethernet, Wi-Fi, USB, Bluetooth, DisplayPort |
| Starting price | $89,999 |
| Pre-order deposit | $999 |
| Lead time | Approximately 3 to 4 months |
Note on payload: detailed product listings give the per-arm payload as 7 kg (15 lb).[13] Some secondary write-ups and the LG CNS coverage instead cite a figure of about 15 kg, with The Korea Herald describing it as a load of "about 15 kilograms with both arms combined," which is consistent with roughly 7 kg per arm.[9]
Vega starts at $89,999, which undercuts most full-size humanoid robots aimed at research customers.[1][13] Pre-orders require a nonrefundable deposit of $999, and the published lead time is approximately three to four months from order confirmation.[13] The pricing has been a recurring point in trade coverage, which frames Vega as one of the lower-cost dual-arm humanoid platforms available to labs and integrators.[1][10]
Vega's most distinctive feature is its foldable torso and arms. The platform compacts from a 171 cm standing height to roughly 66 cm, letting operators move the robot through standard doorways, transport it in a passenger vehicle, and store it under workbenches when not in use. When unfolded, the torso extends so that the end effectors can reach 2.2 meters above the floor, which covers most overhead shelves and conveyor lines used in warehouse and retail settings.[1][13]
The omni-directional wheeled base sidesteps the locomotion challenges that bipedal humanoids still face on real factory floors. Wheels are mechanically simpler, more stable under heavy upper-body loads, and easier to control with classical motion planners. The trade-off is that Vega cannot climb stairs or step over obstacles, so deployments are aimed at flat indoor environments rather than unstructured outdoor work.[10]
Each arm ends in a five-fingered hand. The default configuration uses a 12 DOF hand for fine manipulation; customers can opt for a simpler 6 DOF hand for tasks that require less articulation. The hands are aimed at general-purpose grasping across thousands of household and industrial objects, with Dexmate publicly claiming a 99 percent success rate on the manipulation benchmarks it uses internally. The company emphasizes that it does not try to mimic the appearance of human hands; instead, it targets what Chen and Qin call "effective degrees of freedom," optimizing joint count and placement for downstream task performance.[3][4]
Vega's perception stack is built around several complementary sensors:
| Sensor | Purpose |
|---|---|
| RGB-D cameras | 3D scene understanding, object detection, manipulation feedback |
| LiDAR | Long-range mapping, navigation, obstacle avoidance |
| IMUs | Orientation, balance, motion tracking |
| Ultrasonic | Short-range proximity for safe operation around humans |
| Force/torque | Contact-aware control during grasping and assembly |
The onboard computer pairs an Intel x86 CPU for general-purpose tasks with an NVIDIA Jetson AGX Orin module, available in 32 GB or 64 GB configurations, for real-time AI inference and motion planning. Optional integration with a large language model allows Vega to be tasked using natural language, and the robot ships with a Python software development kit that lets researchers and integrators drop the platform into existing pipelines.[1][13]
Dexmate's stated strategy is to tightly couple custom hardware with data-driven learning so that the two co-evolve.[3]
Chen and Qin describe their approach as engineering the hardware from "day one" with downstream AI training in mind. Joint layouts, actuator stiffness, and sensor placement are chosen for what they make easier to learn, not just what they enable mechanically. The hands combine rigid skeletal cores with deformable outer surfaces to widen grip tolerance and improve tactile feedback, traits that matter more for imitation learning and reinforcement learning than for scripted behaviors.[3]
Dexmate calls its data strategy an "exponential data engine." Rather than relying solely on linear human teleoperation, the company combines three sources:
| Data source | Method |
|---|---|
| Simulation | Reinforcement learning and classical control in NVIDIA Isaac Sim and Isaac Lab |
| Real-world teleoperation | Vision Pro and Quest headsets, custom gloves, and exoskeletons |
| Video learning | Imitation learning from human demonstration video |
Chen has noted publicly that roughly 90 percent of raw reinforcement-learning rollouts are "effectively useless" because of stochastic exploration, so the team prioritizes data quality and curation over raw volume. Its vision-language-action models are built on top of NVIDIA's Isaac GR00T foundation models, which provide pretrained backbones for embodied AI.[3][4]
The technical lineage of the company traces directly to two well-known robot-learning groups: Pulkit Agrawal's lab at MIT CSAIL, which produced influential work on dexterous in-hand manipulation with reinforcement learning, and Xiaolong Wang's group at UCSD, which has published widely on dexterous manipulation, sim-to-real transfer, and learning from human video. Dexmate's founding team co-authored or led several of those papers before spinning out the company.[5][6]
Dexmate targets customers across industries facing labor shortages and high manipulation demands:
| Sector | Example tasks |
|---|---|
| Manufacturing | Assembly, machine tending, kitting, quality inspection |
| Logistics and warehousing | Picking, packing, palletizing, sorting, material transport |
| Retail | Shelf stocking, inventory scanning, restocking |
| Food service | Wok flipping, ingredient handling, cleaning |
| Research | University and corporate robot-learning labs |
| Hazardous environments | Operations in spaces unsafe for human workers |
In its launch materials and interviews, the company has emphasized warehouse and logistics use cases first, citing industry data showing that labor accounts for roughly two thirds of warehouse operating budgets in the United States. Several public companies in manufacturing, logistics, and retail were running Vega pilots by the time of the ProMat 2025 show.[3][8]
Vega's pricing has attracted attention because it undercuts most full-size humanoid robots aimed at research customers. Coverage in trade outlets such as Humanoid.guide and Robotic Gizmos has highlighted the foldable design and dual high-payload arms as standout features in its price tier, while The Korea Herald has described Dexmate's hardware as a research hardware platform used by several robotics intelligence developers.[1][9][13]
Industry observers have also placed Vega in the broader context of the wheeled-base versus bipedal-base debate in humanoid robotics. Wheeled designs sacrifice some flexibility, including stair climbing, in exchange for stability, cost, and battery life, which fits Dexmate's pitch that practical deployment matters more than form-factor purity.[10]