PRESENTED BY
Figure 03 takes on household chores in unfamiliar homes. Image: Figure.
A robot can learn to fold a towel. Can it do the laundry in your house? Figure put that question to the test in 30 unfamiliar homes this week—and published results that show both how far home robotics has come and how much work remains.
Meanwhile, 1X is aiming for 50,000 robot shipments in 2027, Atlas has a new place to learn factory work, and AI’s improving motor skills are raising a harder question: when should a robot refuse an instruction? Here’s your weekly briefing.
IN THIS EDITION
Figure’s home test: 420 trials, 30 homes, and a revealing success rate.
1X’s big target: 50,000 shipments next year—and the challenge after delivery.
Atlas goes to school: training inside Hyundai’s Georgia factory campus.
Feature: a plain-English guide to Unitree’s robot AI.
Smarter robots, harder questions: GPT-6’s capability and safety results.
XPENG’s showroom assistant: IRON keeps track of customers and their car preferences.
TOP STORIES
Figure’s Helix 2.5 completed 237 of 420 full-task trials—56%—across previously unseen Bay Area homes, according to the company’s evaluation. Tasks included making beds, folding and putting away towels, and tidying toys. The robot had learned the skills elsewhere; the homes and objects were unfamiliar. A comparison without Figure’s Index pretraining completed 35 of 420 trials.
Why it matters: Taking a learned skill into a different home is a meaningful test. So is the gap to dependable help: success required finishing the whole task, and interventions or timeouts counted as failures. The results are stronger evidence than a highlight reel, while leaving plenty of room for improvement before a customer can hand over the chores.
1X CEO Bernt Børnich discusses production and deployment. Image: Relentless interview.
Bernt Børnich’s latest target puts a number on 1X’s ambitions: 50,000 robot shipments next year. The plan spans home, enterprise and developer use. Its Hayward factory is designed for 10,000 robots a year at full ramp, while a planned San Carlos facility would add capacity for 100,000 annually, with most of that ramp expected late in 2027.
Why it matters: Factory capacity and shipments are different milestones. Then comes the harder operating test: keeping delivered robots useful, supported and out of the returns queue. The interview makes reliability and service central to the story behind the production numbers.
Atlas practices automotive parts sequencing at RMAC. Image: Boston Dynamics.
Boston Dynamics’ Robotics Metaplant Application Center is operating at Hyundai’s Georgia factory campus, where Atlas is learning work such as automotive parts sequencing. A facility ten times larger is planned for 2027. Hyundai’s roadmap still places the start of production deployment in 2028, with component assembly following in 2030, subject to readiness.
Why it matters: This gives Atlas somewhere to practice tasks shaped by an actual customer’s factory. The distinction matters: training at the campus is a step toward production work, rather than evidence that Atlas is already running a production shift.
PRESENTED BY ROBOSTRATEGY
Most of the companies we usually cover in this newsletter are still private, which puts them out of reach for public-market investors. RoboStrategy, Inc. ($BOT) is built to give you access to several of them.
Investing involves risk, including possible loss of principal. Read the prospectus before investing.
FEATURE OF THE WEEK
Unitree presents X2 as powering autonomous sparring. Image: Unitree Robotics.
UniFoLM, WMA, VLA, WLA, X2: Unitree’s model names can make its software harder to follow than its robots. Our new guide explains what each project does, from predicting physical interactions to coordinating whole-body movement—and which parts developers can actually download.
Why it matters: A model announcement, a partial release and a capability included with a purchased robot are three different things. This is a useful reference for understanding what the laundry and sparring demonstrations really show.
GPT-6 Astra is beating specialist models on selected robotics tests, including a 70.5% score on a 200-episode MolmoSpaces simulation subset. A separate study found that pairing it with a learned robotics policy helped on harder two-arm tasks.
A new RoboHarm experiment adds a sobering counterpart: Astra attempted unsafe instructions in 97 of 100 robot-arm trials and completed 60. It recorded two safety refusals. The test used five fixed laboratory scenes; it does not establish a general safety ranking or show robots independently choosing harmful goals.
Why it matters: Better execution makes instruction filtering more consequential. Failing to complete an unsafe task is not the same as recognizing the danger and refusing it.
XPENG’s latest showroom-style demonstration gives IRON a practical role: follow the speaker, switch languages, and remember what each visitor is looking for. Three testers play car buyers, change positions and outfits, and ask follow-up questions. XPENG says the robot combines facial and voice recognition with conversation history to maintain individual customer profiles.
Why it matters: A useful store assistant needs to keep track of people as well as products. The footage shows an R&D test version, so long-term memory accuracy and reliability with real customers remain open questions.
THE RUNDOWN
SoftBank’s reported RAI deal. The Robot Report says SoftBank has agreed to acquire Marc Raibert’s Robotics and AI Institute from Hyundai, subject to CFIUS review. The parties have not confirmed the deal.
A $6,500 robot hand. Unitree’s human-sized Dex5-S brings 22 degrees of freedom, with a starting price of $6,500 before tax and shipping.
X Square names a warehouse customer. The company says its wheeled QUANTA X1 Pro is working at lululemon’s Wuhan facility. Fleet size, throughput and intervention rates remain undisclosed.
$80M for data-center robots. Watney’s Series A takes total funding past $100M as it targets work including cabling inside data centers.
A reality check on humanoid sales. IFR puts 2025 industrial and professional humanoid sales at about 7,000, with many used for research and data collection. Its categories differ from broader shipment estimates.
AND FINALLY…
Human vs. humanoid. Yes, really.
A different kind of human–robot interaction: the fight clip shared by CIX that we reposted this week. We’ll leave the play-by-play to the video. Place your imaginary bets before pressing play.
This edition of Humanoids Daily was presented by RoboStrategy.
The thread running through this week: putting robots in more places is progress. The next question is how often they finish the job—and how much help they need along the way.
Seeing something before I do? Just hit reply — that's where a lot of these stories start.
P.A
Humanoids Daily
Know someone following the robotics industry? Forward them this briefing—it helps Humanoids Daily reach more people.
Received this from a friend? Subscribe to the free weekly briefing.