TrueSeeker · Verified claim report Case ce489bab02 · 2026-08-17

§ Claim under review · Mixed

"A 77-pound (35 kg) Unitree G1 humanoid robot pulled a 3,086-pound (1,400 kg) car in a demonstration developed with the Beijing Academy of Artificial Intelligence." Accompanying post text: "CHINA'S HUMANOID ROBOT PULLS 3,086-POUND CAR IN STUNNING DISPLAY OF POWER" / "35kg pull 1400kg" / "Developed alongside the Beijing Academy of Artificial Intelligence, the Unitree G1 accomplished the task using high-torque actuators, advanced balance algorithms, and a crouched stance."

Circulating claim, as submitted.

Verdict

Mostly accurate

Confidence

High
§

Summary

This one is real but oversold. A 35 kg Unitree G1 robot did pull a car reported at 1,400 kg, and the video came from the Beijing Academy of Artificial Intelligence, a genuine research institute, which released it in October 2025 alongside a published technical paper on a robot control system called THOR. The catch is the framing. Pulling a car that is already rolling on a smooth, flat floor takes only about 1 to 2% of its weight in force, and the researchers' own paper measures the robot's peak pull at about 168 newtons, which is less than half the robot's own body weight. So the "35kg pulls 1400kg" line makes it sound like a 40 to 1 strength feat when the real achievement is balance and traction under sustained load, which is still genuinely difficult engineering. Two smaller errors: the G1 is an off-the-shelf Unitree robot and was not co-developed with BAAI, which only supplied the software, and the clip is roughly ten months old but is being posted as new. It also remains unclear whether the robot was fully autonomous or receiving direction commands from an operator, and a driver was seated in the car during the test.

§

The readings

key figures from the evidence
167.7 N

peak pulling force of the robot, backward

35 kg

weight of the Unitree G1 robot

1,400 kg

reported mass of the car pulled

§

Why this verdict

Every headline fact in the claim checks out against primary sources: a 35 kg Unitree G1 did pull a car reported at 1,400 kg, the demonstration was released by BAAI, and there is a genuine technical preprint behind it with measured results. This is not fabricated or doctored content. The verdict falls short of fully accurate because the post's framing, particularly the "35kg pull 1400kg" hook and the "display of power" headline, omits the decisive physical qualifier that pulling a rolling car on flat ground requires roughly 1 to 2% of its weight in force, and the paper's own measured peak pull is about 167.7 N, less than half the robot's body weight. The caption additionally misstates that the G1 was developed alongside BAAI when BAAI supplied only the control software, and it presents ten-month-old material as current. Confidence is High because the originating institution, the first-party release, and the quantified technical paper were all located and directly compared against the claim.
§

Evidence

The event is real and originates from an identifiable institution. The Beijing Academy of Artificial Intelligence shared a 21-second clip on social media to demonstrate its humanoid control system called THOR, short for Towards Human-level Whole-Body Reaction.

The robot in the footage is a Unitree G1 from Hangzhou-based Unitree, weighing 35 kg, and the white sedan is stated at 1,400 kg.

BAAI posted the footage on its official X account on 27 October 2025 with the message that it was "Pushing the boundaries of humanoids with THOR."

The underlying research is a published preprint. Thor is described as a humanoid framework for whole-body reactions in contact-rich environments, using a force-adaptive torso-tilt reward function to encourage human-like responses during force-interaction tasks.

It uses a reinforcement learning architecture that decouples the upper body, waist, and lower body, and was deployed on the Unitree G1.

Critically, the paper quantifies what the robot can actually pull. The robot achieves a peak pulling force of 167.7 N, approximately 48% of the G1's body weight, when moving backward and 145.5 N when moving forward, improvements of 68.9% and 74.7% over the best-performing baseline. Thor is also capable of pulling a loaded rack at 130 N and opening a fire door with one hand at 60 N.

The paper's headline real-world demonstrations are listed as opening a fire door requiring 60 N, pulling a loaded rack requiring 130 N, pushing a wheelchair with a 60 kg load requiring 40 N, and wiping a blackboard.

Contemporary reporting also recorded the conditions and the criticism. The footage shows the humanoid making visible effort to pull the car out of a parking area over a flat surface, and after a brief struggle for traction the robot regained control by crouching slightly.

Reporting noted that the force needed to move the car is relatively low because the vehicle is on a smooth, flat surface where friction and rolling resistance are low.

Some commenters questioned the authenticity, with one writing that "1400kg means nothing" because the real pulling force required depends on the coefficient of friction between the rolling wheels and the floor. A driver was seated inside the car during the test, prompting questions about whether it was carried out without human intervention.

§

Findings

What's accurate 8

  • The event happened and is documented by a first-party institutional release, not an anonymous viral repost.
  • The robot is a Unitree G1 and it does weigh approximately 35 kg (77 lb).
  • The Beijing Academy of Artificial Intelligence is genuinely involved. It developed and released the THOR control system used in the demonstration.
  • BAAI is a real, well-established research institution, not an invented organization.
  • There is a real, publicly available technical preprint behind the work, arXiv:2510.26280, with quantified force measurements and comparison against named baselines.
  • The car figure of 1,400 kg is consistently reported across all coverage and traces to the original BAAI framing.
  • The "crouched stance" detail in the caption corresponds to a real technical element of the work, the force-adaptive torso-tilt reward that makes the robot lean to increase applied force.
  • The description of a controlled environment is accurate and was acknowledged in the original coverage.

What's misleading 6

  • Omitted qualifier, the most significant issue. The post's central hook, "35kg pull 1400kg," invites readers to infer that the robot exerted force proportional to a 1,400 kg load, a 40 to 1 ratio. The paper's own measurement is a peak pull of about 167.7 N, roughly 17 kgf, or about 48% of the robot's own body weight. Moving a wheeled car on a flat, smooth floor requires overcoming rolling resistance, which is on the order of 1 to 2% of the car's weight, not its full weight. The impressive engineering achievement is stability and traction under sustained load, not raw strength approaching a ton.
  • Misattributed development. The caption says the Unitree G1 was "developed alongside the Beijing Academy of Artificial Intelligence." This is inaccurate. The G1 is a commercially available Unitree robot. BAAI's contribution is the control software layer, THOR, trained and deployed onto existing hardware.
  • Marketing framing of hardware as the achievement. The caption credits "high-torque actuators" as a driver of the result. The actuators are standard Unitree hardware, unchanged by this research. The paper's claimed contribution is explicitly the reinforcement learning control architecture and reward design, and its gains are measured against other software controllers running on the same robot.
  • Date and context mismatch. The demonstration was released on 27 October 2025. This post carries a publication date of 16 August 2026, roughly ten months later, and presents the material with no date, framed as current news. Recycled content presented as new.
  • Loaded language. "STUNNING DISPLAY OF POWER" and "showcasing impressive strength" apply a strength frame to a result the underlying paper frames as stability and whole-body reaction under contact-rich load.
  • Speculative extension stated confidently. The claims about industrial, warehouse logistics, and disaster response applications are aspirations echoed from the paper's motivation section, not demonstrated outcomes. A parking-garage demonstration does not evidence deployment readiness.

? What's uncertain 5

  • Whether the car-pulling demonstration specifically appears in the Thor preprint with measured figures. The paper's listed real-world demonstrations are the fire door, loaded rack, wheelchair, and blackboard. The car pull appears to be a promotional video accompanying the paper's release rather than a quantified experiment within it. I could not confirm this either way, and search capacity was exhausted before I could retrieve the full paper text.
  • Degree of autonomy. It is not established whether the robot was operating fully autonomously or receiving high-level velocity or direction commands from a human operator. Reinforcement-learning locomotion controllers of this type commonly take operator-issued velocity commands while executing balance and force control autonomously, but I could not verify the specific setup here. This was a live question in the original comment thread.
  • Whether the car was in neutral, and whether the driver applied any pedal input. Reporting notes a driver was present for safety, and one blog states the car was in neutral, but I found no first-party statement confirming the drivetrain state or that no assistive input occurred.
  • The exact 1,400 kg figure. It is uniformly reported but I did not locate a first-party BAAI statement specifying the vehicle model or a weighed mass.
  • Conflicting pull-force figure. One blog reports the robot achieved "up to 95 kg (209 lbs) of measured pull force." This is roughly 932 N and is irreconcilable with the paper's stated 167.7 N peak. I could not verify the origin of the 95 kg figure and treat it as unreliable.
Distortion flags misattribution
§

Sources

7 of 7 linked to records
[1]

Li, Shi, Hu, Hu, Wang, Wang, Luo, "Thor: Towards Human-Level Whole-Body Reactions for Intense Contact-Rich Environments," arXiv:2510.26280 (Oct 2025)

primary preprint by the research team behind the demo
https://arxiv.org/abs/2510.26280 ↗
[2]

BAAI official X account (@BAAIBeijing), post of 27 October 2025 releasing the video

primary first-party release by the originating institution
https://x.com/BAAIBeijing ↗
[3]

alphaXiv technical overview of arXiv:2510.26280

secondary paper-derived summary with reported figures
https://www.alphaxiv.org/abs/2510.26280v1 ↗
[4]

Interesting Engineering, 28 October 2025

secondary technology journalism citing the BAAI post directly
https://interestingengineering.com/innovation/video-unitree-humanoid-robot-pulls-car ↗
[5]

Digital Trends, 28 October 2025

secondary technology journalism
https://www.digitaltrends.com/computing/humanoid-robot-pulls-car-with-ease/ ↗
[6]

VnExpress International, 29 October 2025

secondary general news aggregation
https://e.vnexpress.net/news/tech/tech-news/china-s-35-kg-humanoid-robot-pulls-1-4-ton-car-4957664.html ↗
[7]

mikekalil.com blog, 28 October 2025

tertiary individual blog, contains a pull-force figure that conflicts with the paper
https://mikekalil.com/blog/baai-thor-unitree-humanoid-robot-pulls-car/ ↗
How links are chosen. A source is linked only when the address comes from the investigation's own retrieval or from a registry lookup (PubMed, Crossref) that matches the citation's title and year. Author lists shown as registry-verified come from the registry record, not from the report text. Citations that cannot be matched are labeled, never guessed.
This is one case on the record See the full case, browse the archive, and search every checked claim on TrueSeeker Open on trueseeker.com →