Unitree Go2 Review: Which Tier Actually Does What You Want
Four Go2 tiers, and only the top one gives SDK and ROS 2. Real runtimes, what Air gives up, and two separate security problems worth isolating the robot for.
The Go2 is the robot that made quadrupeds a consumer product. It walks well, it costs less than a laptop at the entry tier, and it is the reason a generation of students has hands-on time with a legged robot at all.
It also has four tiers, and the difference between them is not a spec bump. It is whether you can write code for the machine. That is the same trap the Unitree G1 sets on the humanoid side, and buyers fall into it the same way.
Companion reading: how to choose a robot dog / quadruped, legged and quadruped robot hardware, robot cybersecurity, ROS 2, and edge AI and robot compute.
Table of contents
- Key takeaways
- The four tiers, and what each one withholds
- Runtime and the physical machine
- The software boundary
- Two security problems, not one
- How it compares
- Who should buy which tier
- Verdict
- Frequently asked questions
- Changelog
The four tiers, and what each one withholds
| Tier | Relative price | What it is for |
|---|---|---|
| Air | 1x (the entry point) | Walking, following, consumer demos |
| Pro | ~2x | The sensible consumer buy: better sensing and compute headroom |
| X | ~3x | Between Pro and EDU |
| EDU | ~7x and up, quote-only | Full SDK, ROS 2, Jetson Orin, actuator-level control |
Those multiples held through the mid-2026 price cuts even as the absolute numbers dropped, which makes the ratio the more useful thing to plan against.
The rule of thumb that saves the most money: buy Pro rather than Air unless you have read exactly what Air gives up, and buy EDU if anyone will write code. Air is a capable walking robot and a poor development platform. The middle tiers are the consumer sweet spot. EDU is a different category of purchase with a quote attached.
Runtime and the physical machine
The Go2 is a compact 12-degree-of-freedom quadruped, roughly 76 x 31 x 20 cm crouched, and about 15 kg with the battery.
Runtime is where the tiers separate physically:
- Consumer tiers: 8,000 mAh, roughly 1 to 2 hours
- EDU: 15,000 mAh, roughly 2 to 4 hours
Both figures assume ordinary walking. Continuous high-power gaits, cold weather and heavy sensor payloads all pull them down. As with any legged robot, the energy goes into staying upright, not into travelling, so idle time is not free either.
The software boundary
This is the part worth understanding before the purchase order.
EDU includes full SDK access with ROS 2 support and a Jetson Orin for onboard compute, which is enough to run neural network policies in real time. That combination is why EDU units end up in robot-learning labs: you can train a policy, deploy it on-device, and iterate.
Below EDU, you are working with the app and the higher-level interfaces. The specific thing people discover late: the "talk to your robot" demos that circulate need an LLM wired in through the SDK, which means EDU. A conversational Go2 is not a consumer-tier feature.
For the surrounding stack see ROS 2, and for what a Jetson-class module does and does not buy you, edge AI and robot compute.
Two security problems, not one
Most Go2 coverage mentions neither. Both belong in a deployment decision.
1. The unauthenticated DDS interface on EDU. EDU firmware exposes a programming interface without authentication. Anyone on the same network can reach it, which in practice means an EDU unit can be rooted by anyone sharing its network. This is unlikely to be "fixed" in the usual sense, because open programmatic access is the tier's entire value proposition. Treat it as a design property: EDU units belong on an isolated VLAN, not on a campus or office network.
2. UniPwn. The Bluetooth Low Energy vulnerability disclosed on 20 September 2025 by Andreas Makris and Kevin Finisterre covers the Go2 as well as the B2, G1 and H1. Hardcoded cryptographic keys and an authentication bypass give root from within wireless range, and the exploit is wormable between nearby Unitree robots. As of May 2026 researchers reported no patch and the same authentication architecture still in production firmware.
Neither is a reason not to buy one. Both are reasons to put it on its own network segment and to think twice before running a fleet of them in a public venue. Our robot cybersecurity guide covers the general posture.
How it compares
| Go2 | Boston Dynamics Spot | Unitree B2 | Deep Robotics X30 | |
|---|---|---|---|---|
| Entry cost | Lowest in the category | Highest | Mid-high | Mid-high |
| Programmability | EDU tier only | Full SDK, enterprise | Yes | Yes |
| Payload | A few kg | ~14 kg | Higher than Go2 | Industrial |
| Runtime | 1-2 h consumer, 2-4 h EDU | ~90 min | Longer | Longer |
| Sealing | Light | IP54 | Higher | Industrial |
| Best fit | Research, teaching, prototyping | Industrial inspection at scale | Heavier field work | Industrial inspection |
The Go2's advantage is not capability, it is access. It is the quadruped a student, a lab with a small budget or a startup prototyping a gait can actually own. For what an enterprise inspection deployment costs by comparison, see the Boston Dynamics Spot review.
Who should buy which tier
Air: you want a walking robot to demonstrate, film or play with, and you will never write code for it.
Pro or X: the default consumer choice. Better sensing and headroom, still not a development platform.
EDU: mandatory if anyone will write control code, deploy a learned policy, use ROS 2, or wire in an LLM. Budget the isolated network alongside it.
None of them if you need an IP-rated machine for a real industrial environment, or payload beyond a few kilograms. That is Spot, B2 or X30 territory.
Verdict
The Go2 is the most important quadruped in education and research, for the same reason the G1 matters in humanoids: it is the one people can actually buy. The hardware is good, the tier ladder is honest once you understand it, and the mid-2026 price cuts made EDU meaningfully more reachable for labs.
Buy the right tier the first time, because the gap between "walking robot" and "programmable robot" is the whole purchase. Put EDU units on their own VLAN and keep them off any network you care about, because between the unauthenticated DDS interface and an unpatched wormable BLE exploit, this is a robot that assumes a friendly network and does not get one in the real world.
Frequently asked questions
How many Unitree Go2 versions are there? Four: Air, Pro, X and EDU. Air is the entry point, Pro and X are the consumer tiers, and EDU is the quote-only research configuration that costs roughly seven times the entry tier and up.
Which Go2 do I need to program the robot? EDU. Full SDK access with ROS 2 support and the Jetson Orin compute module are EDU features. The consumer tiers give you the app and higher-level interfaces, not actuator-level control.
How long does the Go2 battery last? Consumer tiers carry an 8,000 mAh pack good for roughly 1 to 2 hours. EDU carries 15,000 mAh for roughly 2 to 4 hours. Hard gaits, cold and heavy payloads all reduce that.
Can the Go2 talk to me like the demo videos? Not out of the box on a consumer tier. Conversational behaviour means wiring your own LLM stack through the SDK, and the SDK is an EDU feature.
Is the Unitree Go2 secure? No, and it has two separate problems. EDU firmware exposes an unauthenticated DDS programming interface reachable by anyone on the same network, which is the tier's core feature rather than a bug to be patched. Separately, the UniPwn BLE vulnerability disclosed on 20 September 2025 allows wormable root-level takeover from within wireless range and was still unpatched as of May 2026. Run these robots on an isolated VLAN.
Go2 or Boston Dynamics Spot? Different categories. Go2 is for research, teaching and prototyping at a small fraction of the cost. Spot is an industrial inspection product sold with fleet software and a support contract, and its configured price is orders of magnitude higher.
Is Air worth saving money on? Usually not. The common advice, which matches the tier design, is to buy Pro rather than Air unless you have specifically read what Air gives up in sensing and compute.
Related guides
- Unitree G1 Review: The Cheap Humanoid's Expensive Catch
- Boston Dynamics Spot Review: The Price Everyone Quotes Is Wrong
- Universal Robots Review: The Lineup Is Two Ladders, Not One
- Franka Research 3 Review: The Lab Default, and Its Corporate Asterisk
- Tesla Optimus Review: The Scorecard, Not the Demo Reel
- Figure 03 Review: The Only Humanoid With a Repeat Customer