K-Scale Labs focuses on building open-source humanoid robots, targeting developers and prioritizing accessibility and affordability.
Current industry alternatives are proprietary and expensive; K-Scale Labs aims to democratize general-purpose humanoid robotics by open-sourcing the full technology stack.
The Kbot is a 4’11” humanoid robot developed in five months, featuring a full aluminum body, robust RL-based locomotion control, and extensive sensor integration.
Kbot is positioned as the most affordable developer- and research-grade humanoid robot, priced at $9,000, compared to alternatives like the $40,000 Uni robot.
All hardware (open BOM, CAD, electronics), software, and machine learning models for Kbot are fully open-source, supporting community replication and modification.
Designed for modularity: users can swap end effectors (e.g., hands, grippers), legs, arms, and heads to match application needs or upgrade as technology advances.
Comes with a Python and Rust SDK for easy programming; capable of utilizing state-of-the-art ML frameworks (e.g., Nvidia Isaac Sim, MJX) and running VLMs via cloud.
Continuous hardware and software improvements are planned, with over-the-air updates released weekly.
Uses MIT Cheetah actuators and offers up to 250 TOPS compute power.
Zbot is a compact (1.5 ft) humanoid robot designed as a more affordable, accessible option, originating from a hackathon project.
Runs the same software stack as Kbot, allowing cross-compatibility for apps and locomotion policies.
The Zbot project has gained strong community interest, with about 5,000 Discord members and many 3D-printed units.
Zbot will also be fully open-source and is approaching mass manufacturing.
Open Source Hardware, Software, and ML Stack 07:54
Both robots and all associated code (hardware, ML, interfaces) are fully open-sourced to promote developer engagement and rapid innovation.
The ML stack is organized for autonomy, using a high-level controller (VLA) and RL-based whole-body motion policies, with custom firmware in Rust.
The overarching goal is to make robot programming so accessible that anyone can create reusable applications, fostering an ecosystem akin to an app store.
A GPU-accelerated training library (using MJX) enables rapid locomotion and manipulation policy development; a walking policy trains in 1–2 hours.
The operating system includes a Python/Rust interface and a simulation environment (“KS sim”) with a unified programming API, allowing seamless transitions from simulation to real-world deployment by just changing the IP address.
The training setup abstracts away most complexity, enabling policy development in under 500 lines of code.
The project leverages a large open-source community (active Discord, public bounties, MIT license) for contributions and rapid development.
Regular hackathons, active hiring, and plans for scaling up the team are in place to sustain and grow progress.
Q&A: Technical Details, Use Cases, and Comparisons 14:02
Power: Robots use battery packs offering around two hours of walk time; robots can run while charging via wall plug.
Use cases: Initial users include developers, researchers, and enthusiasts; the ultimate goal is first US consumer humanoid robotics company, with applications in household tasks as AI models mature.
ROS vs custom stack: Chose a custom, Rust-based OS for simplicity and ease of use, citing negative experiences with ROS setup complexity and minimal async communication needs.
AI hardware: Kbot uses Jetson Orin Nano and AGX, with flexible compute options.
Teleoperation: Supports VR headset control with RL-assisted inverse kinematics for intuitive manipulation.
Tesla comparison: Tesla Optimus is more powerful and factory-oriented, priced around $60,000; Kbot aims for broader use at $9,000 pre-mass production, with comparable general-use capabilities.