Open-Source Humanoid Robots in 2025: A Practical Guide

Why Open-Source Humanoids Matter

Until 2023, meaningful humanoid robot research required either a multi-million-dollar custom platform or access to Atlas or ASIMO through a major research partnership. The combination of commodity servo technology, open-source walking controllers, and decreasing manufacturing costs has fundamentally changed this. In 2025, a lab can get started with a humanoid platform for $8,000-70,000 — within the budget of a university robotics lab or a serious startup.

This democratization is not just about access — it is about community velocity. Open-source platforms accumulate software contributions from dozens of research groups simultaneously, which accelerates the software ecosystem faster than any single company can. The open-source humanoid community now numbers in the tens of thousands of active contributors across Discord, GitHub, and HuggingFace, producing walking controllers, manipulation policies, and simulation environments at a pace that no single company matches.

Full Platform Comparison

Unitree G1: The Most Accessible Entry Point

The Unitree G1 ($16,000) is the current reference platform for accessible humanoid research. At roughly 1.3m height and 35kg, it is a full humanoid with 23 DOF, two 7-DOF arms, and walking capability that has been demonstrated at 2m/s on flat ground and over moderate terrain. The Python SDK is well-documented, ROS2 support is available through community packages, and the growing HuggingFace LeRobot community has published walking policies, manipulation pipelines, and teleoperation setups specifically for the G1.

Important caveat: the G1 at $16K is the base configuration with 6-DOF parallel jaw hands. For dexterous manipulation research, you will need to add dexterous hands ($2,000-8,000 additional) and a teleoperation glove system. Budget $20,000-28,000 for a full manipulation research setup. The EDU variant ($27,000) adds 20 additional DOF in the hands (5 per finger x 2 hands) and includes force-torque sensors at the wrists.

Unitree H1: The Research Workhorse

The Unitree H1 ($90,000) is the G1's larger sibling, designed for research groups that need adult-height reach and higher payload. At 180cm and 47kg, the H1 can interact with human-scale environments (countertops, door handles, shelves) without the reach limitations of the shorter G1. Each arm produces up to 40Nm at the shoulder and 25Nm at the elbow, enabling manipulation tasks that exceed the G1's 15Nm shoulder limit.

The H1's walking is more stable than the G1's due to its larger feet and higher center of mass, with published gait controllers achieving 3.3 m/s on flat ground. The tradeoff is cost and risk: at 47kg, a fall during development poses a real safety concern, and replacement parts for the H1 are 2-3x the cost of G1 equivalents.

Berkeley Humanoid

The Berkeley Humanoid (Zhuang et al., 2024) is a completely open-source design from UC Berkeley's Hybrid Robotics Lab. The full hardware design files, CAD models, bill of materials, and walking controller code are published. The design is optimized for RL locomotion research — lighter and faster than G1, with a focus on whole-body control research rather than manipulation.

This is a build-it-yourself platform. The BOM cost is roughly $8,000-12,000 in parts, with significant assembly time (expect 200+ hours for a first build). For labs with mechanical engineering resources, it provides a customizable platform at lower cost than any commercial option. For labs without hardware expertise, the assembly and maintenance overhead is significant. Several UC Berkeley, CMU, and MIT groups have built Berkeley Humanoids, and the community maintains a shared troubleshooting wiki.

Reachy 2 by Pollen Robotics

Reachy 2 ($70,000) occupies the upper end of the open-source humanoid space. It is a commercial product with fully open-source software (Apache 2.0 license) and partially open hardware. Its 17 DOF arms, human-like kinematics, and native ROS2 stack make it particularly well-suited for research on human-robot interaction and bimanual manipulation.

The higher price buys you better mechanical quality, a supported software stack, and a company that will answer support tickets. For labs doing research that will be published in IEEE/ICRA proceedings and needs reliable, reproducible hardware, Reachy 2 is worth the premium over DIY alternatives. Reachy 2's teleoperation SDK is the most polished among open-source humanoids, with VR-based control that requires minimal setup.

Commercial Humanoids: 1X NEO, Apptronik Apollo, Figure 02

Three commercial humanoid programs are worth tracking even though they are not open-source, because they influence the direction of the field:

1X NEO Beta is the most safety-focused humanoid platform. 1X Technologies (Norway) has designed NEO around the principle that a humanoid operating in homes must be intrinsically safe — soft actuators, lightweight construction (30kg), and force-limited joints. The NEO Beta program gives select research partners access to the hardware and a data collection SDK. If your research focuses on safe human-robot interaction, NEO's design philosophy is the most thoughtful in the field.

Apptronik Apollo is positioned as an industrial platform — heavier (73kg), stronger, and designed for warehouse logistics tasks like box moving and pallet loading. Apollo uses a traditional rigid actuator design optimized for payload rather than safety. Access is through a commercial partnership program with Apptronik, which includes deployment support.

Figure 02 has the most advanced dexterous hands of any humanoid platform, with 16 DOF per hand. The Figure + OpenAI partnership has produced impressive language-conditioned manipulation demos. Hardware access is limited to strategic partners (BMW is the anchor deployment partner), but Figure's published results provide useful benchmarks for what current humanoid manipulation can achieve.

Software Ecosystem Comparison

Choosing a Platform: Decision Framework

The right choice depends on your research goals, budget, and team capabilities:

• Locomotion RL research (budget: <$15K): Berkeley Humanoid. Fully open hardware means you can modify the kinematics, add custom sensors, and iterate on mechanical design. The MuJoCo model is accurate enough for sim-to-real locomotion transfer.
• Tabletop manipulation research (budget: $16K-28K): Unitree G1. The largest community, most available tutorials, and best simulation support. Tabletop reach is adequate at 127cm height. Add dexterous hands from SVRC's catalog for manipulation beyond parallel-jaw grasping.
• Human-height manipulation and mobile tasks (budget: $90K): Unitree H1. Full adult-height reach, strongest arms, and whole-body control capable. Necessary for tasks that involve shelves, countertops, door handles, and other human-scale affordances.
• Human-robot interaction and bimanual research (budget: $70K): Reachy 2. Best teleoperation SDK, most natural arm kinematics, and the Apache 2.0 software license means no commercial restrictions on your research outputs.
• Safe in-home deployment research (budget: invitation only): Apply for the 1X NEO Beta program. The soft actuator design makes it the only humanoid platform where contact with humans is by design rather than by accident.

Data Collection Suitability

For teams using humanoids specifically as data collection platforms (collecting demonstrations to train policies on the same or different hardware), the key criteria are teleoperation latency, joint position accuracy, and recording infrastructure:

Getting Started: A Practical Path

The fastest path from zero to a trained manipulation policy on a humanoid:

• Month 1: Buy Unitree G1, set up Isaac Lab simulation, run walking policy in sim to validate environment. Install Python SDK, verify motor control, run basic teleoperation test.
• Month 2: Deploy pre-trained walking policy to hardware, test stability in lab space, add teleoperation setup for arms. Calibrate cameras, set up data recording pipeline (HDF5 or LeRobot format).
• Month 3: Collect 200-300 manipulation demonstrations via teleoperation, train ACT policy, evaluate on simple pick task. Iterate on demonstration quality — operator training makes more difference than algorithm tuning at this stage.
• Month 4+: Scale to harder tasks, introduce object diversity, collect failure-recovery demonstrations. Contribute to LeRobot community datasets to benefit from shared improvements.

SVRC supports this path through the G1 and H1 hardware packages available in our store, the SVRC G1 support program for labs getting started, and our data collection service for teams that need demonstrations collected on humanoid platforms. We maintain both G1 and H1 units at our Mountain View facility and can collect humanoid manipulation datasets on demand ($2,500 pilot / $8,000 campaign).

Related Reading

• Unitree G1 Review
• Mobile Manipulation Survey
• ALOHA Robot Guide
• Sim-to-Real Transfer
• VLA Models Explained
• Data Services
• Robot Leasing
• Hardware Catalog
• Glossary