Unitree G1 Complete Guide
The Unitree G1 is the most important humanoid robot of 2025-2026, not because it is the most capable, but because it made humanoid robotics accessible. At $13,500 for the base model, it is the first humanoid that universities, small startups, and individual researchers can actually afford to buy, break, and experiment with.
Before the G1, humanoid robot research required $200,000+ hardware or simulation-only work. The G1 changed the economics of who can do physical AI research. This guide covers everything you need to know before buying one.
Core specifications
| Spec | Base model | EDU model |
|---|---|---|
| Price | $13,500, excluding tax and shipping | Contact sales |
| Height | 132 cm | 132 cm |
| Weight | ~35 kg (with battery) | ~35 kg |
| Degrees of freedom | 23 | Up to 43 |
| Hands | Basic grippers | Dex3-1 dexterous hands (3-finger) |
| Battery | 9,000 mAh, quick-release | Same |
| Runtime | ~2 hours | ~2 hours |
| Folded dimensions | 690 x 450 x 300 mm | Same |
| Joint motors | 23 | Up to 43 |
| Development support | SDK (Python, C++) | Full SDK + ROS 2 + LeRobot |
Critical distinction: The $13,500 base model from Unitree’s shop states “does not support secondary development.” If you want to write custom code and control the robot programmatically, you need the EDU configuration. The base model walks, gestures, and demonstrates, but you cannot program it.
For most readers of this site, the EDU model is what you actually want. The base model is for showrooms and demonstrations, not research.
DOF configurations explained
The G1 comes in multiple configurations with different joint counts:
| Configuration | DOF | What changes | Use case |
|---|---|---|---|
| Base (23 DOF) | 23 | Standard joints, basic grippers | Demonstrations, exhibitions |
| Research (29 DOF) | 29 | Added wrist articulation | Basic manipulation research |
| Advanced (36 DOF) | 36 | Full arm articulation + waist | Locomotion + manipulation |
| Dexterous (43 DOF) | 43 | Dex3-1 three-finger hands | Dexterous manipulation, VLA research |
Research papers on the G1 typically use 27-36 DOF configurations. The 43 DOF configuration with dexterous hands is the most capable but also the most expensive and mechanically complex.
For academic publications, the 29-36 DOF range is the sweet spot: enough articulation for meaningful manipulation research without the maintenance overhead of the full dexterous hand system.
What the G1 can actually do
Based on published research and documented capabilities (not marketing demos):
Walking and locomotion:
- Walks on flat surfaces reliably
- Handles modest terrain (slight inclines, low obstacles)
- Recovers from minor pushes (dynamic balance)
- Multiple trained walking gaits available
Manipulation (EDU/developer models):
- Grasps objects up to ~1kg (limited by arm payload)
- Teleoperation via motion capture or controller
- Imitation learning policies via LeRobot/GR00T
- Basic pick-and-place in controlled environments
Development platform:
- NVIDIA Isaac integration (GR00T reference workflow uses G1)
- Hugging Face LeRobot support (both 23 and 29 DOF variants)
- Unitree SDK (Python and C++, low-level motor control + high-level locomotion API)
- ROS 2 bridge available
What it cannot do (honestly):
- Heavy lifting (35kg robot, arms are proportionally weak)
- Complex manipulation in unstructured environments (current policy quality is limited)
- Extended operation (2-hour battery, no hot-swap during task)
- Outdoor operation in weather (not weather-sealed)
- Autonomous useful household tasks (no current policy achieves this reliably)
The development ecosystem
Unitree SDK 2
The official SDK provides two levels of control:
Low-level: Direct motor control. Set joint positions, velocities, or torques for each of the 23-43 motors individually. Useful for researchers implementing custom controllers.
High-level: Locomotion API. Command walking direction, speed, and body pose without managing individual joints. Useful for applications where walking is a means, not the research focus.
Available on GitHub: unitreerobotics/unitree_sdk2
NVIDIA Isaac GR00T + G1
NVIDIA provides a complete end-to-end workflow specifically for the G1:
- Record teleoperation demonstrations on the real G1
- Use demonstrations to fine-tune the GR00T N foundation model
- Deploy the fine-tuned policy back on the G1
This is documented in NVIDIA’s Isaac ROS Physical AI reference architecture. The G1 is one of the officially supported robot platforms for GR00T development.
Hugging Face LeRobot + G1
The G1 is supported in LeRobot for:
- Teleoperation (record demonstrations)
- Imitation learning training
- Locomanipulation policy testing
- Both 23 and 29 DOF variants are supported
Unitree maintains an official LeRobot integration: unitreerobotics/unitree_IL_lerobot
NVIDIA Isaac Sim
The G1 has an official digital twin in Isaac Sim for:
- Policy testing before real-hardware deployment
- Synthetic data generation
- Reinforcement learning in parallel environments
Who is buying the G1
Based on published research papers and public deployments:
- Universities: The majority of G1 buyers. Used in robotics labs for locomotion, manipulation, and human-robot interaction research.
- AI companies: As a physical testbed for learned policies developed in simulation.
- Startups: As a prototyping platform before designing custom hardware.
- NVIDIA partners: Through the GR00T program as a reference platform.
The G1 is rarely deployed for “production” work (it is too small, too limited in payload, and has too short a battery life). It is a research and development tool, not a product that performs useful labor.
G1 vs alternatives
| Robot | Price | DOF | Payload | Best for |
|---|---|---|---|---|
| Unitree G1 | $13,500 base; EDU contact sales | 23-43 | ~1kg arms | Research, education, VLA development |
| Unitree R1 EDU | $10,500 | Limited | Minimal | Cheapest programmable entry |
| Unitree H2 | $29,900 | Full-size | ~5kg | Commercial pilot, heavier tasks |
| 1X NEO | $20,000 | Unknown | Unknown | Home use (if it ships) |
The G1’s unique position: it is the cheapest full-featured humanoid with NVIDIA GR00T support, LeRobot integration, and a growing research community. Nothing else combines price + ecosystem at this level.
Buying considerations
Should you buy a G1?
Yes, if:
- You are a robotics lab that needs a physical humanoid for policy testing
- Your budget is $15,000-$50,000 (depending on configuration)
- You want NVIDIA Isaac/GR00T and LeRobot ecosystem support
- You are willing to invest development time (this is not plug-and-play)
No, if:
- You need a robot that performs useful work today (the G1 is a development platform)
- You need heavy payload (look at the H2 for $29,900)
- You want a home companion robot (nothing is ready for this, the G1 included)
- You only need a desktop-scale manipulation platform (an SO-101 at $100-$350 is sufficient)
What to order:
- For demonstrations only: base model ($13,500)
- For software development: EDU with 29 DOF (contact sales)
- For dexterous manipulation research: EDU with 43 DOF + Dex3-1 hands (contact sales)
Shipping: Ships from China, 4-8 weeks delivery. International shipping adds $500-$1,500 depending on destination. Import duties apply (check your country’s tariff on industrial robots/educational equipment).
FAQ
Can I program the $13,500 base model?
No. Unitree’s shop explicitly states the base model “does not support secondary development.” You need the EDU edition for programmable control. The base model is for demonstrations and exhibitions only.
How long does the battery last?
Approximately 2 hours under normal walking operation. Less if running intensive manipulation tasks. The battery is quick-release, so you can swap in a charged battery without powering down.
Is the G1 supported by NVIDIA Isaac GR00T?
Yes. NVIDIA provides an official reference workflow for the G1 including teleoperation recording, GR00T model fine-tuning, and deployment. The G1 is one of the primary supported platforms.
Can the G1 work with Hugging Face LeRobot?
Yes. Both 23 and 29 DOF configurations are officially supported. Unitree maintains an integration repo (unitree_IL_lerobot) for imitation learning with the G1.
Is 23 DOF enough for research?
Depends on your research. For locomotion research (walking, balance, terrain): 23 DOF is sufficient. For manipulation research (grasping, object interaction): you need at least 29 DOF, preferably 36-43 for dexterous work.
How does the G1 compare to the H2?
The H2 ($29,900) is full-size (180cm, 75kg), with higher payload and stronger actuators. The G1 (132cm, 35kg) is compact and cheaper but limited in what it can physically accomplish. Choose the H2 if you need a robot that can do real physical work; the G1 if you need an affordable research platform.
Sources
- Unitree G1 product page — shop.unitree.com — $13,500 price, “does not support secondary development” disclaimer
- Unitree G1 specs page — unitree.com — 23–43 DOF configurations, Dex3-1 hands, technical specifications
- Unitree G1 in LeRobot — Hugging Face docs — Official LeRobot support for 23 and 29 DOF variants
- LeRobot v0.5.0 release — Hugging Face blog — Full Unitree G1 humanoid support (whole-body control models)
- GR00T End-to-End Workflow for Unitree G1 — NVIDIA docs — GR00T reference workflow: teleoperation, fine-tuning, deployment
- Isaac ROS Physical AI Reference Architecture — NVIDIA — G1 as officially supported GR00T platform
- unitree_IL_lerobot — GitHub (Unitree Robotics) — Official Unitree LeRobot integration for imitation learning