Open Source Humanoid Robots in 2026: A Field Survey
From Open Duck Mini to Zeroth Bot — a survey of the open source humanoid robot projects you can actually build, study, or contribute to today.
On this page
The humanoid robot landscape has changed dramatically in the past two years. While Boston Dynamics and Figure AI dominate headlines, a quieter revolution is happening in university labs, maker spaces, and GitHub repositories around the world.
Here’s a practical survey of the open source humanoid projects that matter right now.
Open Duck Mini
Arguably the most accessible entry point into bipedal robotics today, Open Duck Mini is a 3D-printable, ROS 2-compatible duck-shaped biped that teaches reinforcement learning for locomotion without requiring expensive hardware.
The key insight: by making the robot small (sub-1kg) and inexpensive (< $200 in parts), the team dramatically lowered the iteration cost for sim-to-real transfer experiments. You can crash it hundreds of times without worrying about actuator costs.
What makes it notable:
- Full sim-to-real pipeline using Isaac Gym / MuJoCo
- ROS 2 native, with a clean Python API
- Active community with regular paper releases
Zeroth Bot
Zeroth is taking a different approach: a full-sized humanoid (~1.3m) designed for community contribution from the ground up. The project publishes full mechanical CAD, electronics schematics, and firmware under permissive licenses.
The Zeroth team has been explicit that their goal is creating a platform where AI researchers can test embodied intelligence algorithms without waiting for commercial robot access.
Berkeley Humanoid
Out of UC Berkeley’s HiPeR Lab, the Berkeley Humanoid is a research-grade platform that has produced some of the most-cited recent work on legged locomotion. The full hardware design is open, and the lab regularly releases training code alongside papers.
Recent highlights include their work on expressive whole-body control — getting the robot to maintain balance while performing dexterous upper-body tasks.
What’s Missing
The weakest point across all these platforms is hand and manipulation hardware. Locomotion is largely solved at the research level; dexterous manipulation at human-level capability remains elusive. The open source ecosystem has almost nothing competitive with commercial options like the Allegro or Shadow hand.
This is the gap the community should focus on in 2026.
Getting Involved
The best starting point depends on your background:
- Software / ML background: Start with Open Duck Mini’s sim pipeline. Get comfortable with MuJoCo and ROS 2 before touching hardware.
- Mechanical engineering background: Zeroth’s CAD files are well-documented. Their Discord has active mechanical design discussions.
- Embedded / electronics: Berkeley Humanoid’s electronics stack is the most mature open source offering for full-size platforms.
All of these projects are actively looking for contributors. Start by opening an issue, not a pull request.
