Unitree AS2 in the UK: what we can supply (and why the research still talks about Go2)

If you are evaluating a Unitree quadruped in the UK in 2026, the question is no longer just "Go2 or A2?". The AS2 series occupies the middle ground: about 20 kg in weight, up to about 15 kg of load in continuous motion and over 5 m/s of top speed on the X and EDU versions, IP54 protection on Pro, X and EDU. It suits professional demos, labs and inspection or surveillance where the Go2 starts to feel tight and A2 or B2 would be oversized.

AS2 specs by version (Unitree data)

Source: As2 parameter table on unitree.com. Weight with battery about 20 kg for all versions. Unitree notes that the values are measured in the laboratory and may vary.

This guide combines two levels: (1) Abra's commercial offer with End-User prices excluding VAT; (2) scientific context from recent arXiv preprints on locomotion and deployment of Unitree quadrupeds, because PoC decisions cannot rely on the brochure alone.

Why cite Go2 papers when we talk about the AS2

The AS2 is newer as a commercial SKU exposed in the public price list. The research community (RL, sim-to-real, navigation) keeps publishing heavily on the Unitree Go2 because it is the de facto standard for datasets and benchmarks. Three examples useful for anyone buying an AS2:

• AcL (arXiv:2503.21401) trains fault-tolerant gaits up to four joints and validates them on a real Go2 with one and two joint failures: relevant if your use case involves impacts or faults during inspection.
• MoE + RoboGauge (arXiv:2602.00678) tackles sim-to-real predictability and multi-terrain robustness (snow, sand, stairs) on the Go2, useful as an acceptance-test checklist before go-live.
• Library navigation on a Go2 EDU (arXiv:2606.03340) integrates RTAB-Map and Nav2 in a real library: 100% success in a static scene and 88% with many people, mean metric map error of 3.7 cm, a pattern transferable to the AS2 EDU/Pro in corridors and plant floors.

At Abra we use this literature to define what to measure in a PoC (path success rate, fault recovery, sim-to-real gap), not to promise performance identical to the paper: AS2 hardware, payload and firmware must be validated on your layout.

AS2 models available in the UK (End-User price list)

All the pages are in the AS2 range page. Prices from the public price list, excluding VAT; availability and lead times are confirmed in the quote.

AS2 vs Go2 vs A2: how to choose

• Go2: education, entry-level HRI, plenty of papers and open-source policies (see also Kine2Go, a Go2 kinematic dataset). Ideal if the team needs to reproduce RL papers.
• AS2: the same compact form but with more headroom on payload, IP and perception. Choose it for customer demos, indoor patrol and inspection PoCs where you need a more serious LiDAR/camera (Pro) without moving up to the A2.
• A2 / B2: industrial payloads and protection. The A2 is IP56 (A2 Pro up to IP67 on the main components) with about 25 kg of load in motion; the B2 is IP67 with over 40 kg (A2 page, B2 page). Oversized for a lab, justified outdoors and in demanding plants.

What the research implies for an AS2 PoC in a business

1. Sim-to-real is not free. Work on the Go2 shows that policies trained in Isaac/MuJoCo need domain randomisation and transfer metrics (2602.00678). Budget time for on-site fine-tuning.
2. Dynamic indoor ≠ static map. The library case (2606.03340) reminds us that people and carts degrade the success rate: in a plant, use paths and right-of-way rules.
3. Ground perturbations matter. Training on oscillating surfaces (2510.13488) improves robustness, relevant on walkways and vibrating floors.
4. Vegetated outdoor is a separate failure mode. The ForEnt dataset on forest entrapment (2606.19675): if your case is undergrowth, do not assume an "indoor paper" gait is enough.

Typical software stack

• Ubuntu + ROS 2 Humble on a development PC (EDU models)
• unitree_sdk2 and unitree_ros2 (tested by Unitree on ROS 2 Foxy and Humble). The repository does not yet list the AS2 among the supported robots: compatibility must be checked against the firmware version at the time of ordering.
• Isaac Lab / MuJoCo simulation before deployment (same approach as the cited papers)
• For inspection: camera/thermal payload + logging, see also the inspection guide

Training and lab: Universities and Research. Paper roundup: Unitree papers on arXiv.

Next steps with Abra

1. Open the AS2 hub and identify 1–2 candidate SKUs
2. Request an assessment: find the right robot
3. Define PoC metrics inspired by the papers (path success rate, recovery, hours of autonomy)