# Your Personal Open-Source Humanoid Robot for $8,999 Jingxiang Mo, K-Scale Labs | AI Engineer World's Fair 2025 | 19:26 Source: https://www.youtube.com/watch?v=BS92RdBvI90 Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/your-personal-open-source-humanoid-robot-for-8-999 Published: 2025-07-02 Tags: developer-experience, open-source, reinforcement-learning, robotics ## TL;DR - K-Scale Labs is building open-source humanoid robots for developers, including the $9,000 K-Bot and the smaller Z-Bot. - The robots share a Python and Rust software stack, with simulation, reinforcement-learning tools, and hardware designed for replacement and upgrades. - K-Scale Labs is targeting consumer and developer robotics, while acknowledging that household tasks still depend on better vision-language-action models. ## Summary Jingxiang Mo introduces K-Scale Labs' K-Bot and Z-Bot humanoid robots. The K-Bot is a 4-foot-11-inch, aluminum-bodied robot priced at $9,000 before mass production, while the 1.5-foot Z-Bot provides a cheaper platform with the same software and local-motion stack. K-Scale plans to open-source the bill of materials, CAD, electronics, software, and machine-learning models. Its developer tools include Python and Rust SDKs, a Rust-based operating system, a gRPC digital twin, and GPU-accelerated reinforcement-learning training with MJX. Mo shows a workflow that moves from training in simulation to deployment by changing an IP address. The company is aiming at consumer and developer robotics rather than only factory customers. The talk is also candid about current limits: household tasks such as folding clothes and washing dishes await better vision-language-action systems, and teleoperation remains the practical option today. ## Key ideas ### K-Scale Labs is opening the full humanoid stack to developers [00:01](https://www.youtube.com/watch?v=BS92RdBvI90&t=1s) Jingxiang Mo says K-Scale Labs builds humanoid robots across hardware, software, and machine-learning models. Its goal is general-purpose robotics that more people can use, rather than technology controlled by a small number of expensive, proprietary companies. The team has about 15 people in Palo Alto. The company plans to open-source the complete stack, including the bill of materials, CAD designs, electronics, PCBs, software, and trained models. Mo frames this as a way for developers to replicate the robot and build applications on top of it. The company also runs hackathons and uses public bounties to extend what its small team can deliver. ### The K-Bot is a modular developer robot priced around $9,000 [01:20](https://www.youtube.com/watch?v=BS92RdBvI90&t=80s) The K-Bot is a 4-foot-11-inch humanoid with an aluminum body, sensor coverage, and a reinforcement-learning controller for local motion. Mo says it can take kicks and other damage during demonstrations. The robot is designed to cost $9,000, compared with roughly $40,000 for the next-cheapest option he mentions, and with the price of some industrial robot arms. Its arms can use five or seven degrees of freedom. Developers can replace the hands with parallel grippers or other end effectors, and can replace legs, arms, or the head as the hardware changes. The head can also be upgraded when better compute becomes available. ### The Z-Bot carries the same software ideas into a smaller platform [06:18](https://www.youtube.com/watch?v=BS92RdBvI90&t=378s) For people who cannot spend $9,000, K-Scale offers the Z-Bot, a 1.5-foot humanoid that began as a hackathon project. The project gained attention on Twitter and WeChat, and K-Scale is moving it toward mass manufacturing. Mo says it uses the same local-motion and software stack as the K-Bot, so an application developed for one robot can be moved to the other. The Z-Bot also works with the company's simulators. K-Scale was inspired by Google DeepMind's robot-soccer research, and an earlier 3D-printed version attracted about 5,000 people to the company's Discord, with a few hundred people building one. ### K-Scale wants high-level models and reinforcement learning to work together [08:05](https://www.youtube.com/watch?v=BS92RdBvI90&t=485s) K-Scale's autonomy plan uses a high-level vision-language-action controller together with a reinforcement-learning whole-body local-motion policy. Mo says the company is working on both parts, as well as firmware and a robot software architecture written in Rust. The intended result is a robot that developers can program with ordinary Python applications and share with other users. Those applications could target specific tasks and be reused across robots. K-Scale is also working on integrating and fine-tuning generalist policies from projects such as Pi0 and NVIDIA research, so developers can try different models on the same hardware and infrastructure. ### The SDK and operating system avoid the setup burden Mo associates with ROS [09:32](https://www.youtube.com/watch?v=BS92RdBvI90&t=572s) K-Scale provides Python and Rust interfaces for its robots, with installation through a package rather than a large robotics middleware setup. Mo says the operating system combines a software framework with a Python interface, and lets a developer connect to a robot by IP address. The company chose not to use ROS because its robots do not currently have many sensors that require constant asynchronous communication. Mo also says he had difficult experiences setting up ROS 1 and ROS 2 environments. His preferred workflow is to open the robot, turn it on, and begin programming from a computer. ### The digital twin lets developers move from simulation to hardware by changing the IP [10:07](https://www.youtube.com/watch?v=BS92RdBvI90&t=607s) K-Scale's KS Sim digital twin uses the same gRPC interface as the physical robot. A developer can prototype in simulation without risking the hardware, then change the IP address to send the same control program to the real robot. Mo presents this as a way to shorten the path from an experiment to deployment. The simulator can evaluate whether a trained policy walks, stands, or performs a manipulation task. The software and simulation tools are open source, and K-Scale says developers can begin programming through its package and simulation environment. ### MJX makes local-motion policy training short and reproducible [11:06](https://www.youtube.com/watch?v=BS92RdBvI90&t=666s) Mo describes a training workflow that starts by cloning a repository and running a Python training command. The walking example contains about 500 lines of training code. K-Scale uses MJX for GPU-accelerated simulation and says a walking policy takes roughly one to two hours to train on a local GPU. The training runs through millions of simulated task iterations, with reward functions that developers can tune. TensorBoard provides visibility into losses and rewards. After training, the policy can be evaluated in KS Sim. Once the behavior works in simulation, the same policy can be sent to the robot by changing the target IP. ### The company is aiming at consumer robotics while admitting autonomy is still limited [14:55](https://www.youtube.com/watch?v=BS92RdBvI90&t=895s) When asked about factories versus homes, Mo says K-Scale's bet is to become a US consumer humanoid robotics company. He says early buyers include people interested in household tasks, research, and building business applications with the robot. However, he does not claim that the robots can already do chores such as unloading a dishwasher or folding clothes autonomously. For now, teleoperation is the practical option. K-Scale hopes to add navigation and vision-language-action capabilities as the models improve. The K-Bot's battery lasts about two hours in current testing, and it can run while connected to a wall charger. ## Notable quotes - "Our goal is to really solve general purpose robotics for everyone and open sourcing the entire stack to the entire world." (00:58) - "The entire robot is going to be open source." (02:46) - "All you have to do between programming something in simulation and real is by changing the IP address." (10:09) - "I just want a robot I can just buy, open the box, it stands, walks, and then I can just start programming it using my computer." (17:20) ## Tools & references mentioned - K-Bot - Z-Bot - K-Scale Labs - Tesla Optimus - Unitree robots - 1X - Google DeepMind - MJX - KS Sim - ROS 1 - ROS 2 - Pi0 - NVIDIA - TensorBoard - MIT Cheetah actuators - Python - Rust ## Who should watch - You are evaluating affordable humanoid hardware for a research lab and want replaceable parts, open designs, and a developer SDK. - You want to train reinforcement-learning policies in simulation and transfer them to a physical robot without building the full software stack yourself. - You are interested in household robots but need a clear account of what current systems can and cannot do without teleoperation. ## Related talks - [Reachy Mini: the $300 open source robot you can actually hack](https://aietalks.com/talks/reachy-mini-the-300-open-source-robot-you-can-actually-hack) (Andres Marafioti, Hugging Face, 21:16) - [What Is a Humanoid Foundation Model? An Introduction to GR00T N1](https://aietalks.com/talks/what-is-a-humanoid-foundation-model-an-introduction-to-gr00t-n1) (Annika Brundyn & Aastha Jhunjhunwala, NVIDIA, 17:47) - [Robots as Professional Chefs](https://aietalks.com/talks/robots-as-professional-chefs) (Nikhil Abraham, CloudChef, 18:58) - [The Rise of the AI Software Engineer](https://aietalks.com/talks/the-rise-of-the-ai-software-engineer) (Jesse Han, Morph Labs, 03:48) - [Robotics: why now?](https://aietalks.com/talks/robotics-why-now) (Quan Vuong & Jost Tobias Springenberg, Physical Intelligence, 18:07)