Is the Unitree G1 the best research humanoid?

It's the best default — cheapest open-SDK biped with real export support. "Best" depends on your research: for whole-body control get the G1 EDU; for dexterous manipulation the Fourier GR-2 is stronger; for pure price the R1 or PM01 compete directly.

What's the real difference between the standard G1 and the G1 EDU?

Access and DOF. Standard gives high-level walk/pose commands and 23 body DOF; EDU adds low-level joint control, higher DOF (to 43), and usually more compute and sensing. If your work is the low-level controller, you need EDU — and several times the budget.

Can any of these do autonomous tasks out of the box?

No. All are development platforms: they balance, walk, and accept teleoperation, but autonomy (perception, manipulation policies, safe untethered operation) is work you do. Treat demos as capability ceilings under ideal conditions, not shipped features.

Do research humanoids come with dexterous hands?

Usually not, or only simple grippers. Hands are a separate purchase and a large one — a 16–20 DOF dexterous hand is a meaningful fraction of a standard humanoid's price. Budget for it explicitly; see dexterous hands.

Can I buy a single unit for a lab?

Yes — MOQ is 1 unit across all of these, with lead times of 4–12 weeks depending on model and configuration. Send the exact configuration and destination via the RFQ form for a landed-cost quote.

Why are some prices POA instead of a number?

The GR-2, A2 and higher G1 EDU configs are quoted per project because the configuration (DOF, hands, compute, sensing) drives the price more than the base chassis does. Pin the exact config on the proforma invoice before comparing quotes.

Unitree G1 vs the Field: Research Humanoids Compared

What "research humanoid" actually buys you

Before the spec table, calibrate expectations. Every platform below ships as a development system: a walking, balancing bipedal chassis with an SDK, teleoperation support, and some canned locomotion. None of them arrive able to autonomously do useful work in your environment. You supply the perception stack, the policies, and the integration time.

Concretely, a research humanoid gives you:

A balancing biped you can command via SDK (joint-level or high-level walk/pose).
Teleoperation for data collection (the dominant use today — record demonstrations, then train policies).
Reference locomotion controllers you can run or replace.

It does not give you: robust manipulation autonomy, safe untethered operation around people out of the box, or a maintenance-free duty cycle. Budget engineering months, not weekends. If your goal is a demo rather than a research substrate, that changes which unit makes sense — see the verdicts at the end.

The field, side by side

All figures below are drawn from our catalog fact sheet. Prices are single-unit, EXW China unless a US list price is noted, and are indicative as of 2026.

Model	Height / weight	DOF	Hands included	Price band (single unit)	Lead time
Unitree G1	1.32 m · ≈35 kg	23 (EDU to 43)	Three-finger standard^†; Dex5 optional	$15–50k (from ¥99,000 ≈$14k CN · $16,000 US list · EDU to ≈$74k)	4–8 wks
Unitree R1	1.21 m · ≈25 kg	26	Configurable	$5–15k (¥39,900 · $5,900 US list)	4–8 wks
Unitree H1	≈1.8 m · ≈47 kg	— (full-size biped)	Optional	$50k+ (≈$90,000 class · confirm config)	6–12 wks
Fourier GR-2	1.75 m · 63 kg	53 (incl. 12-DOF hands)	Yes, 12-DOF dexterous	POA (advanced research / healthcare class)	6–12 wks
AgiBot Yuanzheng A2	1.69 m · ≈69 kg	40+	Platform-integrated	$50k+ (reseller-reported ≈$100–190k by variant)	6–12 wks
EngineAI PM01	1.38 m · 40 kg	24 (320° waist)	Optional	$5–15k ($12,000, business & education)	4–8 wks

A few readings of the table:

Price spread is an order of magnitude. A standard G1 or PM01 sits near $12–16k; a fully-loaded G1 EDU pushes to ≈$74k; the A2 class runs $100k+ (reseller-reported). What you pay for at the top end is DOF, payload, runtime and integration support — not "a better research humanoid" in the abstract.
DOF is not a scoreboard. The GR-2's 53 DOF includes two 12-DOF hands; a bare-hands G1 at 23 DOF is a different tool, not a worse one. Count body DOF and hand DOF separately when you compare.
"Hands included" is where budgets die. Most of these ship with simple grippers or no hands. Dexterous manipulation is a separate line item — see dexterous hands, where a Unitree Dex5-1 (20 DOF, 16 active) or Inspire RH56 adds materially to the bill.

The decision that matters: G1 standard vs G1 EDU

For most labs choosing the G1, the standard-vs-EDU fork outweighs the choice between brands. They are not the same robot with a price difference.

Standard G1 ships with a higher-level SDK: you command walking, posture and pre-baked skills. Body DOF is 23. This is enough for teleop data collection, HRI studies, and running Unitree's locomotion. It is not enough if your research is the low-level controller itself.
G1 EDU unlocks the additions researchers usually assume are standard: higher DOF configurations (to 43), low-level joint torque/position access, and typically the compute and sensing options (LiDAR, more powerful onboard compute). This is the tier for whole-body control, RL locomotion, and anything where you replace Unitree's stack with your own.

Practical implication: if a grant proposal says "we will develop novel whole-body controllers," you need EDU, and your real budget runs into the tens of thousands — up to ≈$74k for a loaded EDU — not the $16k headline. Confirm the exact SDK access level and DOF count on the proforma invoice before paying — "EDU" covers several configurations. Get the low-level API scope in writing.

The same axis applies across the field. The PM01's business/education version and the Fourier and AgiBot research platforms all vary in how much low-level access you get. Ask every vendor the same three questions:

Do I get joint-level (torque/position) control, or only high-level walk/pose commands?
Is the SDK open (documented, English, sample code) and does it run on my OS/ROS version?
What sensing and compute is onboard vs. add-on?

Where each unit fits

Unitree G1 (standard) — the default entry point. Cheapest credible path to a balancing biped with a real SDK and an actual overseas support team. Right for teleop data collection, HRI, and coursework. Wrong if you need low-level control or dexterous hands out of the box — spec the EDU and hands, and re-price.

Unitree G1 EDU — the research workhorse. When your object of study is the controller, the policy, or whole-body manipulation, the EDU's low-level access and higher DOF are the reason to buy Unitree over a cheaper biped. Budget for the configured price, not the base.

Unitree R1 — the price floor. At ¥39,900 / $5,900 US list it is the cheapest way to get a bipedal platform onto the lab floor. Smaller (1.21 m, ≈25 kg) and newer; strong for teaching, early experimentation, and fleets where per-unit cost dominates. Verify SDK depth before assuming parity with the G1.

EngineAI PM01 — the G1's closest price-band rival. $12,000, 1.38 m, 40 kg, 24 DOF with a 320° waist — comparable size class to the G1 at a fixed, transparent price. Worth quoting head-to-head with a standard G1 if list-price predictability matters and you don't need Unitree's ecosystem.

Unitree H1 — buy it for dynamics. Full-size (≈1.8 m) and fast (3.3+ m/s demonstrated), at the $50k+ / ≈$90k class. This is a locomotion and dynamic-motion platform, not a manipulation one. Overkill and over-budget for most manipulation-focused labs.

Fourier GR-2 — buy it for hands and payload. 53 DOF including 12-DOF dexterous hands, 3 kg single-arm payload, 1.75 m / 63 kg. POA, research/healthcare class. The choice when dexterous manipulation and a human-scale arm are the point and the standard-humanoid hand story isn't enough.

AgiBot Yuanzheng A2 — buy it for an integrated platform with runtime. 40+ DOF, ≈2 h continuous runtime, 1.69 m / ≈69 kg, reseller-reported ≈$100–190k. A heavier, more turnkey-leaning platform at project pricing; makes sense when you want the vendor's full stack rather than a bare chassis to build on.

Total cost is more than the sticker

The headline price is the start of the bill, not the end. For any of these, plan for:

Hands, if manipulation matters — a separate quote (see dexterous hands).
Compute and sensing on standard tiers — LiDAR (e.g. sub-$1k class units), onboard GPU.
Spare parts — actuators and fingers fail; get them priced at PO time.
Import, shipping and battery paperwork — lithium packs need a UN 38.3 summary; see how to buy robots from China for the full path.

Compare vendors on landed, configured cost — not EXW headline prices. A $12,000 PM01 and a "$16,000" G1 can invert once you add the hands and access tier one of them needs and the other doesn't.