Your Personal Open-Source Humanoid Robot for $8,999 — Jingxiang Mo, K-Scale Labs

Key Concepts

• Humanoid Robots
• Open Source Robotics
• RL (Reinforcement Learning) Controller
• VR Teleoperation
• Modularity
• Python/Rust SDK
• VLM (Vision Language Model)
• OTA (Over-The-Air) Updates
• GPU Accelerated Robot Learning
• MJX
• KS Sim (Kscale Simulation)
• B2C (Business-to-Consumer) Robotics

Kscale Labs Overview

Kscale Labs, based in Palo Alto, is a team of 15 people focused on building open-source humanoid robots, encompassing hardware, software, and machine learning models, specifically for developers. Their goal is to democratize general-purpose robotics by open-sourcing the entire technology stack. They are launching robots in the coming months.

Kbot: The Affordable Humanoid Robot

Overview

The Kbot is a 4'11" humanoid robot developed in five months. It features a full aluminum body, an RL controller for locomotion, and comprehensive sensor capabilities. It is positioned as one of the cheapest humanoid robots on the market, priced at $9,000, and is available for pre-order with delivery expected by October.

Key Features

• Open Source: The entire robot design, including BOM (Bill of Materials), CAD files, electronics, PCB layouts, software, and machine learning models, will be fully open-sourced.
• Modularity: The robot is designed for easy modification and upgrades. End effectors can be swapped, allowing for different hand designs or grippers. New hardware updates, such as legs, arms, or heads with improved compute (e.g., new Nvidia chips), can be easily integrated.
• Software Development Kit (SDK): A Python/Rust SDK is available for immediate programming, supporting state-of-the-art ML algorithms.
• Locomotion and AI: The robot can use Nvidia's Isaac Sim for training locomotion policies and MJX for accelerated compute. It can also interface with VLMs via the cloud.
• Degrees of Freedom (DOF): The robot comes standard with five DOF arms, but can be upgraded to seven DOF arms.
• Over-The-Air (OTA) Updates: Continuous model and software improvements will be rolled out via OTA updates.

Technical Specifications

• Uses MIT Cheetah actuators.
• Includes IMUs, audio modules, displays, and cameras.
• Offers up to 250 TOPS of compute.

Development Timeline

The project started in October of the previous year, progressing from a 3D-printable "Kscale Stompy" prototype to a mass-manufacturable design.

Zbot: The Compact Humanoid Robot

Overview

The Zbot is a 1.5-foot humanoid robot originating from a hackathon project. Due to its popularity on social media platforms like Twitter and WeChat, it is being brought to mass manufacturing.

Key Features

• Shared Software Stack: The Zbot runs the same locomotion and software stack as the Kbot, allowing for code reuse between the two robots.
• Inspiration: Inspired by Google DeepMind's robot soccer research.
• Community Engagement: The 3D-printable version has a strong community following, with over 5,000 Discord members and hundreds of printed robots.
• Open Source: The Zbot will also be fully open-sourced.

Development Timeline

The Zbot project also started in November and is rapidly progressing towards mass manufacturing.

Machine Learning and Software Stack

Core Goal

The primary objective is to make the Kbot fully autonomous, utilizing a dual-policy approach with a high-level VLM controller and an RL-based whole-body locomotion policy.

Software Architecture

The team has developed its own firmware and software architecture in Rust to power the robots. The goal is to create an "app store" ecosystem where developers can create and share Python applications for specific robot use cases.

Developer Tools

• GPU Accelerated Robot Learning Library: An open-sourced library for GPU-accelerated locomotion and manipulation training using MJX.
• Operating System: A software framework with a Python interface, simplifying robot programming compared to ROS 1 and ROS 2. Installation is as simple as pip install kscale.
• KS Sim: A digital twin simulation environment with the same gRPC interface as the real robot, allowing for rapid prototyping and testing without risking damage to the physical robot.

RL Training and Deployment Example

Training Setup

1. Clone the repository.
2. Run python -m train.py. The training code is abstracted into approximately 500 lines for walking.
3. Train the policy using a local GPU with MJX acceleration. Training a walking policy takes approximately one to two hours.
4. Monitor loss and reward functions using TensorBoard.

Evaluation and Deployment

1. Evaluate the trained policy in KS Sim using k inverse.
2. If the simulation results are satisfactory, change the IP address to deploy the policy to the real robot.

Community and Hiring

Kscale Labs actively engages with its open-source community, which includes over 5,000 Discord members. They offer open bounties and run bi-monthly hackathons. They are currently hiring electrical, firmware, and ML engineers.

Q&A Highlights

• Power Source: The Kbot is battery-powered, with a battery pack located behind a panel. Battery life is approximately two hours for walking, but it can also be powered via a wall plug.
• Target Use Cases: Kscale Labs is targeting the B2C market, aiming to be the first US consumer robotics company. Initial customers have expressed interest in using the robots for household tasks and research.
• Household Chores: While current capabilities are limited to teleoperation, the long-term goal is to enable tasks like folding clothes and dishwashing as VLM capabilities improve.
• Reasons for Not Using ROS: The team opted not to use ROS due to its complexity and the fact that their robot doesn't require complex sensor communication. They prioritize ease of use and a model-based approach.
• AI Accelerator: The Kbot uses Jetson Nano and AGX for AI acceleration, with different compute options available.
• VR Teleoperation: VR headsets can be used for teleoperation, with a pseudo-IK system trained using RL for hand gesture and movement replication.
• Comparison to Tesla Optimus: While Tesla's Optimus is more powerful mechanically, Kscale Labs believes their robot offers similar use cases at a significantly lower price point ($9,000 vs. an estimated $60,000+ for Optimus).

Synthesis/Conclusion

Kscale Labs is making significant strides in the field of open-source humanoid robotics with their Kbot and Zbot platforms. By focusing on affordability, modularity, and ease of use, they aim to democratize access to robotics technology and foster a vibrant developer ecosystem. Their open-source approach, combined with a strong community and rapid development cycles, positions them as a key player in the emerging consumer robotics market.