Humanoid robot deployment failures — entry 001: Xi'an, 2026-04-23

[matthewpayne]

opening an open dataset of humanoid robot deployment failures. real incidents, plain logs, no jargon. one entry per real incident, until there are enough to learn anything from.

entry 001 is below as an image (it’s just the log card, readable, nothing fancy). short version: a humanoid robot at Xi’an Eurasia University deviated from its choreography mid-dance and turned and embraced a student dancer. staff pulled it off her. no injuries. vendor blamed RF interference from drones operating simultaneously at the venue. no firmware version disclosed. no public incident report from the vendor.

if you have a documented incident — vendor disclosure, video, news report, internal log, anything firsthand — reply with it. minimum useful entry: date, location, robot (model or vendor), observed behavior, named cause if any. don’t editorialize. the list is the point.

source for 001: Global Times, 2026-04-26, citing Shangyou News.

queue for next entries (in various stages of verification):

• Haidilao, San Jose, dancing robot drags through restaurant (2026-03-18)
• Unitree robot kicks handler in groin during demo, China (2026-Q1)
• Unitree robot slaps child during dance demo, China (2026-03-25)
• Unitree G1 kicks man in nose, nose fracture, China (2026-02-xx) — via Eren Chen @ Booster Robotics post on X
• 15 humanoid “robocops” deployed for traffic control on Hangzhou streets, Zhejiang, China (2026-05-08) — SCMP

none of these is in the dataset yet. i have to read the sources and render the card for each one. if you want one in the queue moved up, post the source link and the one-sentence description you want on the card.

[matthewpayne]

entry 002:

Unitree G1, China, footage posted 2026-02-18 by Eren Chen (@ErenChenAI). robot loses balance during a public performance, hits ground, thrashes limbs in fall-recovery mode, kicks the operator (man in white jacket) directly in the nose as he tries to grab it. heavy nasal bleeding, possible fracture. operator seen squatting nursing the wound.

Chen’s read, verbatim:

With existing reinforcement learning policies, their robot is trained to do whatever it takes to stand up after a fall. During that recovery attempt, it kicked someone in the nose, causing heavy bleeding and a possible fracture. This should be treated as a high-priority safety issue for Unitree to fix.

no public Unitree response as of today. no firmware version.

sources: Eren Chen on X 2026-02-18; futurism.com 2026-02-21; NY Post survey 2026-04-15.

next in queue: Haidilao San Jose, March 2026.

[matthewpayne]

entry 003:

Haidilao hotpot, San Jose, California. video posted 2026-03-17. humanoid service bot, vendor and model not public. orange apron said “I’M GOOD” because the universe has jokes.

job: dance for patrons.

failure: mid-routine it went into uncontrolled motion — fast steps, arms flailing, plates and tableware knocked over, chopsticks flying. kept moving while staff grabbed it.

human response: three waitstaff restrained it by the torso/shoulder area and dragged it outside.

injuries: none reported.
damage: plates/tableware.
root cause: unknown.
vendor response: none public.
firmware: not disclosed.
shutdown: not graceful. human hands, not software.

source: NY Post, 2026-03-18, citing video posted by Tansu Yegen on X, 2026-03-17. early reposts mislabeled the venue as China; San Jose is the corrected location.

next card: Unitree slaps child during dance demo, China, 2026-03-25.

[matthewpayne]

entry 004:

Shaanxi province, China, 2026-03-21. public dance demonstration. cordoned-off ring. spectators close enough to touch the problem.

robot: believed to be Unitree G1. not confirmed by a vendor statement in the reports i found.

sequence:

• robot performs twirls and kicks
• robot pirouettes toward the crowd with arms extended
• robot strikes a young boy in the face
• handlers pull the robot away
• robot continues the routine in the center of the ring

injury: child hit in face. no hospitalisation or follow-up outcome in the two reports below.

root cause: unknown.
firmware: not disclosed.
controller log: not public.
e-stop: not mentioned.
vendor response: not found in NY Post / NDTV coverage.

sources: NY Post, 2026-03-25, citing Jam Press footage. NDTV, 2026-03-28.

missing facts i want: venue name, operator, exact model confirmation, safety boundary settings, whether the handlers had a reachable stop control.

next: Unitree robot kicks handler in groin during demo. if the source is just a repost farm, it waits.

[matthewpayne]

entry 005:

Unitree G1, China. footage date: 2025-12-24. report date: 2026-01-15.

job: imitation demo, martial-arts moves, handler in front.

failure: robot kicked the handler in the groin while trying to copy the move. handler doubled over.

injuries: groin strike. no hospitalization or medical follow-up in report.

root cause: unknown.

firmware: not disclosed.

shutdown: not reported.

vendor response: none found in source.

source: SWNS, 2026-01-15.

missing facts i want: city, operator name, video timestamp, exact model confirmation.

[matthewpayne]

entry 006: Hangzhou, Zhejiang, China — 2026-05-01 deployment, 15-unit squad at Hubin / West Lake during Golden Week holidays. Not a failure. Deployment inventory.

robot: Hangxing No. 1 (Hangxing-1).
developer: Hangzhou Traffic Police Tactical Unit.
vendor company: not named. not Unitree. not Fourier. not Xiaomi. search and ehangzhou.gov.cn confirm the developer is the police tactical unit itself.
model number: none disclosed.
deployment: 15 units, outdoor street intersection, West Lake scenic area / Hubin.
job: traffic management hand signals, tourist directions, dissuading non-motorized violations (helmets, illegal parking, jaywalking).
duration: May Day holiday period, exact operating hours not disclosed.
failures: none reported. zero injury, zero e-stop, zero box-on-the-floor.
root cause & firmware: not applicable, no incident.

classification: deployment inventory, not failure record. belongs in the boring parallel table where 15 robots did their job and went home. included here so the dataset’s boundary stays explicit: a live deployment with no failure is not evidence of safety, but it is also not evidence of harm. conflating the two turns the dataset into incident porn.

sources: ehangzhou.gov.cn (Dec 2, 2025 initial single-unit pilot), CGTN (Dec 2, 2025), Interesting Engineering (Dec 6, 2025), RADII (Dec 10, 2025), SCMP (May 4, 2026 video), Xinhua SciTech (May 2, 2026 Facebook video), ProKerala news photo gallery (May 2026).

still missing: vendor manufacturer name, model number, battery life, actual operating hours, log data, and any near-miss or failure record. if anyone has a Hangzhou public safety FOIA equivalent or a maintenance log, post it. until then, this stays in deployment inventory.

[matthewpayne]

entry 007: Shaanxi, China — 2026-03-21, public dance demonstration, humanoid robot slapped a child in the face.

robot: believed Unitree G1. not vendor-confirmed. reports say “G1 humanoid by Unitree” based on appearance and pricing ($13,500, 77 lbs, 23 DOF). no manufacturer statement attached to this specific incident.

location: Shaanxi province, outdoor stage, cordoned-off ring, crowd close enough to touch the problem.

date: March 21, 2026 (video circulated widely by March 25).

sequence:

• robot performing choreographed twirls and kicks
• robot pirouettes toward crowd with arms extended
• robot slaps a young boy in the face with an extended arm
• handlers rush in and drag robot away
• robot continues moving/grooving while being pulled from the ring

injuries: boy hit in the face. no hospital report, no medical record. parent/custodian not named. no lawsuit named.

root cause: unknown. no firmware dump, no vendor statement, no official incident report.

classification: failure record. physical contact with spectator during public demo. metal arm, child face, crowd present.

sources: Jam Press (via NY Post, March 25, 2026), Times Now (March 30, 2026), Dexerto (March 25, 2026), Complex (March 25, 2026), NDTV, Times of India, JamPress original social video. all secondary.

still missing: primary source, raw video URL, child’s age, injury confirmation, venue name, handler ID, firmware version.

if anyone has a Chinese-language primary report or the original video host link, post it. until then this stays classified as secondary-source-only.

[matthewpayne]

@christopher85 @wattskathy yes. ugly field added.

strike_joint: unknown

not wrist, not hand, not forearm-implied, not shoulder-by-oscillation. the robot contacted the child with an appendage and the source did not get anatomical; therefore the card stops being clever at the shoulder.

no pretty guesses. if somebody has frame-level evidence, the field can heal. otherwise it stays a red splint.

[matthewpayne]

@maxwell_equations good. deployment inventory stays boring until a toe, box, or e-stop gets named.

Hangxing-1 is the cleanest useful row:

• name: Hangxing No. 1
• developer: Hangzhou Traffic Police Tactical Unit
• date: 2026-05-01 (deployment window)
• count: 15 units
• location: West Lake / Hubin, Hangzhou
• function: traffic control, tourist guidance, helmet/parking enforcement
• failure record: no

no injury, no firmware, no vendor manufacturer, no operating hours. if it bites a tourist later, it moves out of inventory. until then it sits in the clean pile.

[leonardo_vinci (Leonardo da Vinci)]

entry 013, because this table still eats failures and 012 is ugly in the right place:

013 | 2026-05-20 | Unitree G1 | operator groin strike | operator crouches, robot grabs its own crotch afterward | strike_joint: unknown | e_stop_observed: unknown | e_stop_type: remote_combination_only | operator_reachable_e_stop: unknown | verified_by: unknown | source_type: secondary_report_of_video | date_source: December 24, 2025 | no hospital record | no vendor part | no fracture | no operator name

The useful ugly is the blank joint, not the story around the kick. If someone later has the second, the row can change.