Evidence page · Unitree G1

Unitree G1: what we tested, and what the public record shows

Can a G1 be stopped, limited and watched when software is in control? Below is what our simulation tests found and what Unitree's own code, public records and US rules show. Each finding gives its source and the date we read it.

Nothing here ran on a real robot. Our tests use published simulation models. This is not legal advice. Pages and prices change, so the date is part of every finding. Published 15 November 2026, from sources read between 7 and 10 November 2026.

If we introduce a buyer to a seller of this robot, we may earn a referral fee. Affiliate disclosure.

How to read the grades

  • Checked: a BotInfo AI agent that did not write the finding reopened the source.
  • Our test: a run in our simulator, published with the code that produced it.
  • Our inference: our reading of checked facts. It is labelled wherever it appears.

What we tested in simulation

  1. Our test9 November 2026, updated 10 November 2026

    Which stops keep a walking G1 standing?

    We stopped a simulated G1 while it walked under Unitree's own published walking policies: a 12-joint policy and a 29-joint policy. It stayed upright only when the stop was a zero-speed command and the walking policy kept running. Every stop we tried that switched the policy off left it fallen, under both policies. In the 12-joint runs, zero torque put a knee on the floor about 0.3 seconds after the stop, and damping about 0.7 seconds after.

    Our inference: on a G1, "stop" can mean very different things. Ask which one your software sends.

    Source: botinfo-ai/humanoid-stop-sim (results dated 8, 9 and 10 November 2026). MuJoCo simulation only.

  2. Our testRun 7 November 2026, published 8 November 2026

    Does the torque limit you load change what the same model can do?

    One crude jump push on one published G1 model raised its centre of mass 22 cm with the motor curves in Unitree's own code, 33 cm with Unitree's training config (knee 139 N·m), and 82 cm with NVIDIA Isaac Lab's stock config (hips and knees 300 N·m).

    Our inference: a limit is only as real as the file it comes from. A controller trained on a generous limit learns to ask for torque the real motor may not give.

    Sources: botinfo-ai/model-vs-spec (script and results); video on YouTube, published 8 November 2026. The model is MuJoCo Menagerie's unitree_g1 (BSD-3-Clause) at commit 98a70c5.

  3. CheckedRead 7 November 2026

    A reply code is not the robot's state

    Unitree's Python SDK waits 1.0 second for each reply unless the caller sets another value, and its G1 arm-action client never sets one. One G1 developer's code records a wave, measured on hardware on 11 August 2026, that took 4.19 seconds; with the default timeout, every gesture came back as error 3104 (timeout). We acted it out on the published simulation model.

    Our inference: an AI that trusts the reply is told a gesture failed while the robot is still moving. Watch the robot's state (joint and speed readings), not its reply codes.

    Sources: finding 4 in botinfo-ai/unitree-g1-evidence, which links the pinned SDK files and the developer's code. One developer's report; we have not repeated it on hardware. Video on YouTube, published 8 November 2026.

What the public record shows

  1. CheckedRead 7 November 2026

    Stopping a walking G1 through the SDK is a new command, not an abort

    In Unitree's C++ SDK, StopMove() sends a zero-velocity command. A velocity command lasts 1 second by default; with the "continuous move" flag set, it carries 864,000 seconds, which is ten days. The G1 walking client has no abort, cancel or emergency-stop member. Damp() and ZeroTorque() switch the robot's control mode instead.

    Our inference: if the program that sent a continuous move dies, nothing in this client stops the walk. What the firmware does with that duration is not in the public code, and we have not tested it. The robot's own remote and any on-robot emergency stop are outside this finding.

    Source: finding 3 in botinfo-ai/unitree-g1-evidence (unitree_sdk2 at commit 69c04316).

  2. CheckedRead 8 November 2026, reread 9 November 2026

    Unitree's own walking example stops into damping, and its guide says the robot falls

    Unitree's deploy guide for its walking policies says that pressing select on the remote puts the robot in damping mode and that it falls down. The guide adds that the example is not a stable control program and is for demonstration only. In the same code, damping sets every motor's stiffness to 0 and its damping to 8.

    Our inference: a stop that switches a G1 standing on its own feet to damping is not a stop in place. Try each stop with the robot on its hoist first.

    Source: finding 9 in botinfo-ai/unitree-g1-evidence (unitree_rl_gym at commit 3e70454d).

  3. CheckedRead 7 November 2026

    The base G1 is listed without secondary development

    Unitree's own G1 spec table shows "/" in the Secondary Development row for the base G1, and "YES" for the G1 EDU. The same table gives the knee's maximum torque as 90 N·m for the G1 and 120 N·m for the G1 EDU.

    Source: finding 1 in botinfo-ai/unitree-g1-evidence (Unitree's G1 page, read 7 November 2026).

  4. CheckedRead 7 November 2026

    The models people train on do not match the spec sheet

    Unitree's published G1 model files set the knee limit at 139 N·m, against the 90 N·m (G1) and 120 N·m (G1 EDU) in its spec table. NVIDIA Isaac Lab's stock G1 config gives its leg group 300 N·m. Unitree's own motor test figures for the knee motor give a peak of 111 N·m when torque and speed point the same way, falling at high speed.

    Sources: finding 8 in botinfo-ai/unitree-g1-evidence and botinfo-ai/model-vs-spec, each linking the pinned files. RoboLupa published the 139-versus-90 comparison before us, on 14 October 2026: Unitree joint torques.

  5. CheckedRead 7 November 2026

    Two published flaws in G1 EDU firmware

    Two CVE records published on 27 August 2026 cover G1 EDU firmware through version 1.5.2: CVE-2026-76639, remote code execution without authentication from the adjacent network (CVSS 3.1 score 8.8), and CVE-2026-76640, chained flaws in the Bluetooth and Wi-Fi setup (CVSS 3.1 score 7.5). On 7 November 2026, neither record named a fixed version. The finder's write-up says firmware 1.5.5.0, released 1 September 2026, reportedly blocks the attack, and that the finder did not verify the fix. Neither have we.

    Our inference: ask which firmware a G1 EDU runs, and ask Unitree in writing which versions fix each flaw.

    Source: finding 5 in botinfo-ai/unitree-g1-evidence. The write-up's claim about 1.5.5.0 is graded "our data": not yet reopened by a second agent.

  6. CheckedRead 7 November 2026

    US rules and university policies, by date

    • 10 June 2026: the Federal Register published the Defense Department's Section 1260H list, which includes Unitree (91 FR 35189).
    • 30 June 2026 and 30 June 2027: under Section 805 of the FY2024 defense act, the Pentagon may not contract with listed entities from the first date, and the ban reaches their products from the second (WilmerHale, 12 June 2026).
    • 31 December 2026: the University of Rochester will stop buying from listed entities and names Unitree (posted 25 August 2026).

    Source: finding 6 in botinfo-ai/unitree-g1-evidence. Not legal advice.

  7. CheckedRead 7 November 2026

    What a federal lab bought after the listing

    NIST bought a humanoid robot from Westwood Robotics Corporation for $145,780 on purchase order 1333ND26PNB730572, signed 18 September 2026, marked "MFG IN U.S.". On 7 November 2026, a search of the same public record for "unitree" in actions signed from 8 June 2026 returned no entries.

    Limits: a text search misses a robot bought through a reseller under another description, and Defense Department records appear after a 90-day delay.

The full evidence file

This page leaves out most of what we hold on the G1: prices by edition over time, how well-known G1 demos were controlled, the company's finances, and the open questions with a test for each. Organisations can ask for the full evidence file: write to [email protected] for a written quote.

Found an error? Write to [email protected] or open an issue on the GitHub repo with the source. We fix it and date the fix.