K-Scale Labs
Last reviewed
Sources
20 citations
Review status
Source-backed
Revision
v7 · 2,623 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
v7 · 2,623 words
Add missing citations, update stale details, or suggest a clearer explanation.
| K-Scale Labs | |
|---|---|
| General information | |
| Full name | K-Scale Labs, Inc. |
| Founded | 2024 |
| Founders | Benjamin Bolte, Matthew Freed, Pawel Budzianowski, Willem Van Rijswijk |
| Headquarters | Palo Alto, California, United States |
| Industry | Robotics, Open-source hardware |
| Products | K-Bot, Z-Bot, Stompy (Zeroth-01) |
| Total funding | approximately $4 million to $4.75 million |
| Valuation (Feb 2025) | approximately $50 million |
| YC batch | Winter 2024 (W24) |
| Team size at peak | approximately 10 |
| Status | Defunct (November 2025) |
| Website | kscale.dev (archived) |
K-Scale Labs was an American robotics startup that developed open-source humanoid robots. Based in Palo Alto, California, the company was backed by Y Combinator and positioned itself as a builder of affordable, hackable humanoid platforms for developers, researchers, and hobbyists. K-Scale shut down in November 2025 after failing to secure a lead investor for its Series A, having shipped only two units of its flagship K-Bot before ceasing operations. Upon closing, the company released all remaining proprietary hardware and software under open-source licenses.[1][2][3]
K-Scale is often cited as the most prominent attempt to build an "Android for the real world," a low-cost, broadly accessible humanoid platform. Its failure also became a widely discussed case study in running a hardware company on software-startup economics, especially in a global humanoid market dominated by rapid Chinese price cuts from firms such as Unitree.[4][5]
K-Scale Labs was founded in 2024 by Benjamin Bolte, a former senior engineer at Tesla who wrote the CUDA code for the voxel occupancy network behind the Tesla Optimus perception model, and later worked on Meta's AI research team on robotics foundation models and semantic mapping. Co-founders included Matthew Freed (advanced robotics at Lockheed Martin and General Dynamics); Pawel Budzianowski, the software lead, a machine learning researcher with a Ph.D. in dialogue systems from the University of Cambridge and prior experience as ML lead at PolyAI; and Willem Van Rijswijk, who handled hardware and operations.[6][7]
The name referenced the Kardashev Scale, an astrophysics framework that measures a civilization's technological advancement by energy consumption. Bolte described the mission as "building infrastructure to make humanity into a Type I Kardashev civilization."[8]
K-Scale started in a Palo Alto garage with roughly nine people, joined Y Combinator's Winter 2024 batch (W24), and used the three-month accelerator to ship its first hardware iteration, the Stompy prototype.
K-Scale raised an initial $500,000 in pre-seed from Lombardstreet Ventures and Y Combinator, then approximately $4 million in a seed round at a roughly $50 million valuation in February 2025. Reported backers included GFT Ventures, Pioneer Fund, AI Grant, Lux Capital, Fellows Fund, and angels Nat Friedman and Daniel Gross. Total publicly reported funding sat around $4 million to $4.75 million.[3][5][9]
OpenAI's nonprofit arm donated $250,000 to K-Scale's open-source work, and OpenAI CEO Sam Altman was reported to have pre-ordered a K-Bot. In June 2025, a subsidiary of Chinese vehicle manufacturer Zhejiang Taotao Vehicle Industry (Tao Motors) invested approximately $2 million and committed supply chain support, including manufacturing capacity in Texas. Even with those contributions, K-Scale never raised a Series A.[3][10]
Before K-Bot, K-Scale built an early prototype called Stompy (Zeroth-01), introduced in February 2024. It was a 122 cm tall, fully 3D-printable humanoid with a bill of materials under $10,000. The team produced five hardware iterations over three months during the YC batch. Every part was sized to fit on a standard 256 x 256 mm 3D printer bed, and all CAD files, assembly guides, and bills of materials were made public. Stompy used quasi-direct drive actuators with 6:1 to 8:1 reduction ratios, providing back-drivable, low-inertia joints with peak torques of roughly 3 to 12 Nm. Power came from a 48V 15Ah hot-swappable battery good for over an hour of runtime.[11][12]
Stompy was one of the first credible "garage-built" humanoid robots and validated the thesis that meaningful humanoid research could happen outside well-capitalized industrial labs.[11]
K-Scale opened K-Bot pre-orders in summer 2025 at $8,999 for the first 100 Founder's Edition buyers. Pre-orders eventually totaled over $2 million from more than 100 customers, including engineers at OpenAI, Nvidia, and Amazon. The first units shipped in October 2025.
On November 4, 2025, Bolte sent a letter to investors and customers announcing that K-Scale was winding down. The company had less than one month of runway and roughly $400,000 in cash. Acquisition talks with 1X Technologies and The Bot Company collapsed after both signaled interest in acqui-hiring select engineers rather than acquiring K-Scale outright. Only two K-Bot units shipped before remaining pre-order deposits were refunded and most of the ten-person team was laid off. In the same letter, Bolte announced that all remaining IP would be released under open-source licenses, framed as an attempt to "lay the foundation for future hackers and dreamers."[1][2][13]
A post-mortem from K-Scale's former chief operating officer identified several strategic errors:[14]
In a podcast after the shutdown, Bolte called the decision to launch the expensive K-Bot ahead of the cheaper Z-Bot his single biggest misstep. A VC had told him that 100 K-Bot pre-orders would unlock a $20 million Series A; that did not happen. "I kind of had bet the farm on the K-Bot launch," he said, adding that he hesitated to lead with Z-Bot because he wanted K-Scale taken seriously as a humanoid company rather than a toy maker.[15]
The macro context did not help. When K-Scale launched its $8,999 K-Bot, comparable U.S. humanoid hardware typically cost $50,000 or more. By late 2025, Unitree was selling its G1 humanoid in the U.S. for about $20,000, and Songyan Power had introduced the first sub-10,000 yuan humanoid in China, intensifying a global price war that compressed margins faster than any new American entrant could absorb.[5]
K-Scale developed three hardware platforms, all designed to be fully open in hardware and software with publicly available CAD files, bills of materials, and assembly guides.
| Model | Height | Weight | Price | Target audience | Status |
|---|---|---|---|---|---|
| Stompy (Zeroth-01) | 122 cm | varies | approximately $10,000 (BOM) | Prototype / proof of concept | Discontinued |
| Z-Bot (Zeroth Bot) | approximately 46 cm | small | approximately $999 assembled; $350 DIY kit | Students, hobbyists, educators | Launched on Kickstarter (2025) |
| K-Bot | 140 cm (4 ft 7 in) | 34 kg (77 lb) | $8,999 Founder's Edition | Researchers, developers | Discontinued (2 units shipped) |
The K-Bot was K-Scale's flagship and the product the company staked its survival on. It used commercial off-the-shelf (COTS) components to keep cost roughly 60% below earlier $50,000+ research humanoids while still delivering 20 degrees of freedom and useful payload capacity.
| Specification | Value |
|---|---|
| Height | 140 cm (4 ft 7 in) |
| Weight | 34 kg (77 lb) |
| Degrees of freedom | 20 |
| Payload capacity | 10 kg (22 lb) |
| Peak torque (max joint) | 120 Nm |
| Battery life | up to 4 hours |
| Sensors | RGB camera, stereo vision, ultra-wide FOV camera, far-field microphones, stereo speakers |
| End effector | parallel gripper hands (modular, upgradable) |
| Onboard computer | Raspberry Pi 5 |
| Operating system | K-OS (Rust-based) |
| Price (Founder's Edition) | $8,999 |
| Manufacturing partner | Tao Motors (Texas, U.S.) |
The K-Bot also offered an optional "Full Autonomy" upgrade path: K-Scale committed to deliver free hardware and software improvements until customers' robots could operate with less than one human intervention per day, targeting June 2028 for that milestone.[16]
The Z-Bot, sometimes branded as the Zeroth Bot, was a roughly 46 cm tall desktop-scale humanoid for students, classrooms, and hackers. It came out of a K-Scale hackathon and reached a working prototype in about two months. Buyers could pre-order an assembled unit for around $999 or a DIY kit starting at $350. Despite its size the Z-Bot supported RL-based walking, basic vision and audio, and torque-controlled joints. K-Scale launched it on Kickstarter in early 2025; after the closure Bolte cited the decision to lead with K-Bot instead of Z-Bot as a key strategic mistake.[15][17]
Stompy was the precursor to both later robots and remains historically significant as the first fully open-source, 3D-printable humanoid a small team could assemble in a garage. Several independent developers built their own units after the design dropped.[11]
K-Scale Labs built an unusually broad open-source software stack. Most of it is preserved on the company's GitHub organization, github.com/kscalelabs, which by late 2025 hosted more than 130 repositories spanning firmware, simulation, learning, inference, and teleoperation.[18]
| Component | Purpose | Language | License |
|---|---|---|---|
| K-OS (kos) | Real-time robot OS with hardware abstraction and gRPC APIs | Rust | MIT |
| PyKOS | Python client for K-OS, used for high-level control | Python | MIT |
| K-Sim (ksim) | RL training library for humanoid locomotion and manipulation, built on MuJoCo and JAX | Python | MIT |
| ksim-gym | RL environments and tasks built on K-Sim | Jupyter / Python | MIT |
| KOS-Sim | Open-source simulation backend integrated with K-OS | Rust / Python | MIT |
| kinfer | Model export and on-device inference tool | Rust | MIT |
| EdgeVLA | Multimodal vision-language-action model for on-device execution | Python | open weights |
| firmware | Actuator and IMU firmware stack | Rust | MIT |
| kteleop / kbot_vr_teleop | Teleoperation interfaces, including VR | Python | MIT / Apache-2.0 |
| onshape | Library for programmatic interaction with OnShape CAD | Python | MIT |
K-Scale was unusual in choosing to skip ROS entirely. The team built its control stack in Rust to get deterministic real-time performance and avoid the operational overhead of ROS. K-Sim, the RL library at the heart of the training pipeline, was reported to generate more than 100,000 training samples per second using MuJoCo MJX on JAX, making large sim-to-real policy training tractable on modest hardware.[19]
EdgeVLA, K-Scale's vision-language-action model, ran on-device on the Raspberry Pi 5. It took voice and visual input and emitted action tokens, giving the K-Bot a foundation for natural-language task execution without a cloud round-trip.
| Component | License |
|---|---|
| Hardware designs | CERN-OHL-S-2.0 |
| K-OS, K-Sim, kbot, zeroth-bot, kinfer | MIT |
| Selected tooling | Apache-2.0 or GPL v3 |
At shutdown, Bolte committed the company's remaining IP to non-commercial open-source licenses so the work could continue outside the company. Hardware was released under the CERN Open Hardware License Strongly Reciprocal v2 (CERN-OHL-S-2.0), which requires derivative hardware to remain open. Most software stayed on MIT or compatible licenses.[2]
At its peak the K-Scale Discord had more than 2,000 active members. The company ran public hackathons under the K-Hacks brand, including K-Hacks 0.1, which produced the Zeroth-01 design. K-Scale also published Hugging Face datasets and trained weights, making it one of the few humanoid efforts that exposed both training code and trained policies publicly.
Open source was central to Bolte's argument for why a humanoid company should exist at all. On the YC Hacker News launch thread he wrote that he would not personally trust a humanoid robot in his home if its software were not auditable. The company's materials framed humanoids as something to be "shared by everyone" rather than controlled by a handful of well-capitalized labs.
Every hardware revision shipped with full CAD source and a public bill of materials. Firmware and the operating system were open from day one. When the company closed, the open-source posture meant the shutdown did not erase the work: independent developers, university labs, and the successor company Gradient Robotics could continue iterating without any IP unwind.
Not everyone was convinced. Hacker News commenters compared the K-Bot to lower-cost academic platforms like Berkeley Humanoid Lite (around $5,000) and raised concerns about military weaponization, given that open hardware can be modified for any purpose. Bolte engaged directly, arguing that an open ecosystem was safer than a closed one in the long run.[20]
After the closure, former Head of Product and Engineering Jingxiang Mo (JX Mo) established Gradient Robotics in Menlo Park, California. Mo described it as having "inherited the K-Scale spirit, mission, and core engineering team," framed the company as the "open-source Unitree for America," and closed a pre-seed round in September 2025. Bolte was not part of the new company, telling interviewers he was "mostly decompressing."[13][15]
K-Scale's run became a reference point in the debate about whether a Western company can compete on price with Chinese humanoid manufacturers without massive upfront capital. The company shipped only two flagship units, yet its software stack, design files, and post-mortem material continue to circulate as a starter kit for anyone building an open humanoid platform. It is also cited as a cautionary case on hardware-versus-software economics: pre-orders did not equal product-market fit, open source did not translate into margin, and a "bet the farm" launch left no fallback when the Series A did not close.