K-Scale Labs: Open-source humanoid robots, built for developers

6 hours ago 1

We're accelerating the timeline to a world with billions of robots, and making sure they're accessible, auditable, and beneficial to humanity.

Join the RL training challenge

Products

General-purpose humanoid robots for developers, hobbyists, and researchers

We're building a robotics stack that lets you deploy physical AI in the real world

Our software, hardware, and machine learning stack is seemlessly integrated, allowing you to focus on building applications instead of installing packages.

Owning every layer—from metal to model—lets us move faster, integrate more deeply, and open-source each breakthrough for the whole world to build on.

Demos

Experience our humanoid robots in action

Application layer

Create, deploy, and share robot apps, tools, and policies—built by you and the community.

Robot Apps Store

A single hub to install apps, customize behaviours, train new skills, and teleoperate your robot.

K-Lang

Control your robot with a neural interpretation domain-specific language (DSL).

ML layer

Open-source library for GPU-accelerated robot learning and zero-shot sim-to-real transfer

K-Sim

High-performance reinforcement learning framework optimized for training humanoid robot locomotion, manipulation, and real-world deployment.

High versatility for tasks such as walking, dancing, household organization, and even cooking.

K-VLA

A generalist policy using large-scale robot data with novel network architecture to enable the most capable and dexterous robot

Locally run and capable of integrating with other VLAs such as Pi0.5 and Gr00t.

OS layer

Rust-based OS to run policies on the real robot, or evaluate in simulation

K-OS

Rust-based, fast, and reliable robot operating system combining hardware, software, and firmware.

Ships with an easy-to-use Python SDK to easily develop robot applications.

K-OS Sim

A digital twin / model evaluator for the K-Scale Operating System (K-OS) using the same gRPC interface as our real robot

Effortlessly test and refine your models in simulation

Hardware layer

Deploy policies, VLA models and applications on robot hardware in real-time

K-Sim

A lightweight, modular framework for developing reinforcement learning policies in simulation and subsequently deploying on physical robots

Made for speed

Train policies in under an hour. Leveraging GPU accelerated libraries like JAX and Mujoco XLA, K-Sim can process over 100,000 samples per second (including training) on an RTX 4090.

Smooth troubleshooting

Quickly pinpoint the root cause of training regressions or improvements. All of the code defining experiment lives neatly in a single file which is automatically logged to TensorBoard.

Optimized for sim-to-real

A trained policy is immediately ready to be exported and deployed on a real robot, saving countless hours otherwise spent on maintaining two sets of infrastructure for training and deployment.

Extensible by design

K-Sim can be easily modified to fit your use-case, instead of the other way around. All of the code is fully open source so you can build on top of it with confidence.

K-SimIssacBraxGazeboDifferentiable physicsGPU-acceleratedTPU-acceleratedCross-platformSingle-file training scriptsOfficially supported hardwareMotion capture priorsStateful (RNNs, Transformers)