LeRobot Humanoid: An Open, Low-Cost, 3D-Printed Humanoid for Robot Learning
TL;DR
- We are releasing LeRobot Humanoid, an open, low-cost humanoid robot project designed for robot learning.
- The current bipedal platform costs around $2,500 in parts, depending on sourcing, shipping, and taxes, and is built from 3D-printed parts, off-the-shelf components, and affordable electronics.
- Unlike many humanoid projects that focus only on a robot model, a controller, or a one-off prototype, LeRobot Humanoid is released as a full stack: hardware, assembly documentation, runtime, identification tools, and training environments.
- The project is aimed at builders and researchers who want a humanoid platform they can build, modify, break, repair, simulate, train, and control.
- The platform is experimental, but already connects the full loop: design exploration → simulation → data collection → identification → training → real-world control.
A humanoid you can build, modify, and learn from
Humanoid robots are becoming an increasingly important target for robot learning. They are general-purpose, human-scale platforms, able to interact with the same objects, tools, and environments as we do. But today, experimenting with humanoids remains difficult: the hardware is often expensive, closed, fragile, or hard to reproduce.
This creates a bottleneck for open robot learning. Even when models, datasets, simulators, and training code are released, the physical platforms used to generate data and validate policies often remain inaccessible. Many people can train policies in simulation, but only a small number of labs can close the loop on real humanoid hardware.
LeRobot Humanoid is designed for a specific niche: an open humanoid platform that is affordable enough to reproduce, simple enough to modify, and complete enough to support real robot-learning experiments.
If you are looking for the most advanced humanoid robot, this is not it.
If you are looking for a humanoid you can build, understand, repair, instrument, simulate, and use for learning experiments, this is the robot we are trying to make.
This is why the release includes more than a robot model or a controller. It includes hardware files, assembly documentation, simulation assets, runtime tools, identification pipelines, and training environments. The goal is to make the full humanoid learning loop accessible, from mechanical design to real-world control.
LeRobot Humanoid is designed as a low-cost platform for real robot-learning experiments.
What we are releasing
LeRobot Humanoid is not a single repository or a single robot model. It is a full-stack release intended to make the platform buildable, inspectable, and useful for robot-learning experiments
The release is organized around five main components:
| Component | What it provides |
|---|---|
| Hardware | A bill of materials, 3D-printable parts, assembly instructions, wiring documentation, and motor setup tools for building the robot. |
| Design tools | A control-oriented design workflow based on simplified robot representations, benchmark tasks, optimal-control evaluation, and design comparison. |
| Runtime | Tools to calibrate and control the robot in simulation and on real hardware, with safety checks and LeRobot integration for data collection. |
| Identification | Dataset replay and simulator parameter fitting tools to reduce the gap between the real robot and simulation. |
| Training zoo | MJLab training environments, including LeRobot Humanoid and other legged robots, to train and evaluate locomotion policies. |
Together, these components cover the main steps of the robot-learning loop: design the platform, build it, simulate it, collect data, identify the model, train policies, and deploy them back to the robot.
The release covers the full loop from hardware design and assembly to simulation, identification, training, and real-world control.
A low-cost, 3D-printed humanoid platform
The first constraint was reproducibility.
A humanoid robot that only works as a one-off prototype can be useful for a demo, but it is much less useful for open robot learning. For this project, we prioritized parts that can be printed, replaced, sourced, and modified by other builders.
The current bipedal platform costs around $2,500 in parts. It is built from a combination of 3D-printed mechanical components, off-the-shelf hardware, and affordable actuators and electronics. The goal was not to minimize cost at all costs, but to find a practical balance between affordability, mechanical performance, and ease of assembly.
This also changes how the robot can be developed. When a structural part breaks, it can be reprinted. When a design choice does not work, it can be modified and tested again. Instead of treating hardware as a fixed artifact, the platform is designed to support fast iteration.
The hardware release therefore includes the files and documentation needed to reproduce the bipedal platform: printable parts, a bill of materials, wiring information, assembly instructions, and motor setup guidelines. The objective is to make the robot something builders can inspect, repair, and adapt to their own experiments.
The current bipedal platform is built from 3D-printed parts, off-the-shelf hardware, and affordable actuators/electronics.
Closing the loop between design, data, and simulation
LeRobot Humanoid is not only a hardware project. It is also an attempt to make the full robot-design loop more reproducible.
Instead of treating mechanical design, simulation, and control as separate steps, the project is organized around a continuous loop. Before building the robot, simplified robot representations are used to evaluate design choices through control benchmarks. After building the robot, real datasets are used to identify simulator parameters and improve sim-to-real consistency.
The design workflow starts from simplified robot representations rather than detailed CAD geometry. These representations make it possible to explore design choices quickly, evaluate them on benchmark motions, and compare candidate mechanisms before committing to hardware.
Once the physical robot exists, the loop continues in the other direction. Real robot logs can be replayed in simulation, and the difference between simulated and measured behavior can be used to identify better simulator parameters.
Training and controlling the robot
Once the robot exists physically, it also needs to exist as a learning target.
LeRobot Humanoid is therefore exposed through simulation assets, training environments, and a runtime stack that can control both simulated and real hardware. The goal is to make the same robot usable across the different stages of the learning workflow: training policies in simulation, testing controllers safely, collecting real-world data, and eventually deploying policies back to the physical platform.
For training, LeRobot Humanoid is available in lerobot-legged-zoo, together with other legged robots. This makes it possible to train and evaluate locomotion policies in MJLab environments, compare behaviors across platforms, and iterate on tasks before running anything on hardware.
On the deployment side, the runtime stack provides a common interface for simulation and real robot control. It handles calibration, state reading, command sending, and safety checks, while keeping the door open for LeRobot integration.
This closes the loop between learning and hardware. The simulator is not only used to train policies, and the real robot is not only used for demos: both are connected through the same runtime and data flow.
Current status and roadmap
LeRobot Humanoid is still an experimental platform, but the current release already provides the foundations needed to build, simulate, control, and improve the robot.
The main released hardware target is the bipedal platform. This is the version currently used for bring-up, calibration, data collection, identification, and early locomotion experiments. The stack is intended for researchers and builders who are comfortable working with robotics hardware, including motor setup, wiring, calibration, and safety procedures.
Running policies on the real robot still requires careful bring-up. Calibration, safety checks, low-gain testing, and a reliable power cutoff are essential before deploying learned controllers. The platform should be treated as an open research robot, not as a finished consumer product.
Current work focuses on improving the robustness of real-world behaviors, closing the sim-to-real gap, and extending the platform toward more complete whole-body capabilities. The first release focuses on the bipedal robot; upper-body integration and more advanced behaviors are part of the roadmap.
Early sim-to-real standing policy on the real robot
From simulation to hardware: an early policy running on the bipedal platfore
Get started
LeRobot Humanoid is released as an open workspace so that builders, researchers, and robot-learning practitioners can explore the platform at different levels.
Start from the main lerobot-humanoid repository for an overview of the full stack.
If you want to build the robot, the lerobot-humanoid-hardware repository contains the bill of materials, printable parts, wiring documentation, assembly steps, and motor setup instructions.
If you want to work on control and deployment, start from lerobot-humanoid-runtime. It lets you test the robot in simulation before connecting real hardware, and provides the tools needed for calibration, state reading, command sending, and safety checks.
If you want to improve sim-to-real consistency, lerobot-humanoid-identification provides tools to replay real datasets in simulation and fit simulator parameters.
If you want to train policies, lerobot-legged-zoo provides MJLab training environments for LeRobot Humanoid and other legged robots.
This release is only a starting point. If you build the robot, train a policy, improve the simulator, or adapt the design, we would love to see what you make.
