Humanoid robots remain out of reach for most people due to their high cost, often priced higher than a car.

Now, a new open-source project called LeRobot Humanoid by Hugging Face aims to change that with a low-cost, 3D-printed design built for learning and experimentation.

While the current version focuses only on robotic legs, the platform offers an affordable entry point into humanoid robotics development.

Priced at around $2,500, the system is designed for hobbyists, researchers, and developers looking to explore advanced robotics without the massive financial barrier typically associated with humanoid machines.

In April 2025, New York-based Hugging Face entered hardware for the first time by acquiring the French startup Pollen Robotics, creators of the open-source humanoid robot Reachy 2.

Open hardware platform

LeRobot is built as a modular bipedal robot platform using 3D-printed mechanical components, off-the-shelf hardware, and affordable actuators and electronics. The current version costs roughly $2,500 in parts, significantly lower than most humanoid systems used in robotics laboratories. Rather than focusing on polished consumer-grade robotics, the platform is designed for experimentation, rapid iteration, and accessibility.

The hardware package includes printable mechanical files, a complete bill of materials, assembly instructions, wiring documentation, and motor setup tools. Structural components can be easily reprinted and replaced, allowing developers to test design modifications without rebuilding the entire system. This approach enables faster hardware iteration and makes the robot more practical for open-source research environments, according to the firm’s website.

The robot currently focuses on lower-body locomotion, functioning primarily as a humanoid biped platform. While upper-body integration and more advanced whole-body manipulation are part of the future roadmap, the existing system already supports standing, walking experiments, calibration, and locomotion policy testing.

The project also introduces a control-oriented design workflow. Instead of beginning with detailed CAD geometry, developers use simplified robot representations to evaluate design concepts through benchmark control tasks and optimal-control simulations. These lightweight models make it easier to compare mechanisms, validate balance strategies, and optimize motion performance before committing to physical hardware production.

Once the robot is assembled, real-world datasets generated from the physical platform can be replayed in simulation to improve model accuracy. This identification pipeline helps reduce the sim-to-real gap by fitting simulator parameters based on actual robot behavior. As a result, simulation environments more accurately reflect real-world hardware performance, improving policy transfer reliability.

Learning control stack

LeRobot is designed as a complete robot-learning platform instead of just a humanoid robot. The project combines hardware, simulation, software tools, and training systems into one ecosystem that supports the full robotics development process.

The platform includes a runtime stack that works with both simulated and real robots. It provides tools for calibration, monitoring robot states, sending commands, and performing safety checks. This allows developers to safely test robot controllers in simulation before using them on physical hardware, reducing the risk of damaging the robot during experiments.

For AI and locomotion research, the robot is integrated into the LeRobot-legged-zoo AI framework, which includes open-source MJLab simulation environments. These environments help researchers train and test reinforcement learning policies for walking and movement tasks. Developers can also compare performance across different legged robot platforms.

According to Hugging Face, a key feature of the project is its sim-to-real workflow. Real-world robot data can be replayed in simulation to identify differences between virtual and physical behavior. Engineers can then adjust simulator settings to improve accuracy and make trained policies transfer more reliably to real hardware.

The release includes hardware files, design tools, runtime software, identification pipelines, and training environments. Although still experimental, the platform aims to make learning for humanoid robots more affordable, open, and accessible to researchers, developers, and robotics enthusiasts.

Recommended Articles