Humanoids daily

Humanoids Daily brings you the latest developments in robotics, with a special focus on humanoid robots and intelligent machines. Newsletter for weekly updates.

[Humanoids daily @humanoidsdaily]

San Mateo County just unanimously voted to draft humanoid robot regulations, and the details are aggressive. The proposed ordinance threatens to completely upend the remote teleoperation business model.

Key mandates from the resolution:

• Robots MUST be supervised at all times by a "trained, on-site human."
• Businesses must pay a "County Automation Impact Fee" to fund worker retraining.
• Annual fees will fund specialized HazMat gear for high-heat lithium-ion battery fires.

Startups launching $30/hr remote cleaning services next door in SF just hit a massive legislative wall.

The regulatory race for physical AI has started.

[Humanoids daily @humanoidsdaily]

This week in humanoid robotics: wild IPO swings, sub-9-second sprints, and a major architectural shift toward in-context learning.

Here is what went down in this week’s Humanoids Daily briefing:

• Unitree’s Market Whiplash: After surging 629% on debut, shares pulled back 45% from their intraday peak. CEO Wang Xingxing gave an unusually candid talk on why factory rollouts remain 2–3 years away.

• Athletic Records Fall: Tiangong Ultra set a blistering 8.86s 100m sprint mark at the World Humanoid Robot Games in Beijing—though braking after the finish line remains an unsolved challenge.

• In-Context Robot Learning: Skild AI (S1) and Generalist AI (GEN-1.5) unveiled models capable of learning complex manipulation tasks from video demonstrations and brief prompts without heavy fine-tuning.

• Industrial Realities: Figure AI proved a 15-foot mezzanine ladder climb driven by real customer workflows, while XPENG raised $900M to scale its IRON humanoid.

• Shipments Surge: Global H1 deliveries topped 22,000 units according to Counterpoint Research, with AGIBOT and Unitree leading the charge.

Read the full edition: https://t.co/B52pXFoI4Z
Subscribe to get our weekly deep-dives straight to your inbox:
https://t.co/CLjb1qy6Pc

[Humanoids daily @humanoidsdaily]

And here:
https://t.co/BHobkIMxpL

[Skild AI @SkildAI]

Introducing S1, our new foundation model that learns from one example.

It can be taught 10-minute long tasks that it has never seen before, from one video prompt without any fine-tuning.

Watch S1 operate in real-time via in-context learning:

[Humanoids daily @humanoidsdaily]

Skild AI's S1 takes a video demonstration as its prompt. No language instruction, no fine-tuning.

Show it a task it has never seen, and it gets 66% of the steps right. Describe that same task in words instead, and you get 9%.

Prompting a robot with video isn't new. Generalist AI showed it earlier this month, but those runs last seconds. Skild claims S1 is the first to work on 10-minute tasks outside the training distribution.

A new task used to mean a hundred hours of teleoperation and a training run. Skild set up a plant potting task at 8:54pm and had the robot running it autonomously 11 minutes later.

[Humanoids daily @humanoidsdaily]

Read more here:
https://t.co/qWaFUqu4Or

[Humanoids daily @humanoidsdaily]

Read more on Humanoids Daily:
https://t.co/3JzGdsHw3j

[Humanoids daily @humanoidsdaily]

🤖 New Humanoid Alert: Meet the P0 from Pink Robotics

Darmstadt-based startup Pink Robotics has emerged from stealth with its first prototype, the P0: a 1.75m, 82kg biped designed for logistics sorting.

Key specs:
▫️ Height: 1.75 m
▫️ Weight: 82 kg
▫️ Degrees of Freedom: 32 DoF
▫️ Payload: 8 kg
▫️ In-house 3D-printed & laser-welded chassis

Founder @ChrisH_elmke is pitching the project as an urgent effort to revive German engineering, voicing sharp frustration with the current domestic status quo following a visit to this year's Hannover Messe.

An industrial pilot with a major supermarket chain is planned next.

[Humanoids daily @humanoidsdaily]

Researchers taught a Unitree G1 humanoid to play tennis using broadcast footage of Federer, Nadal, and Djokovic—and then naturally put it in a polo shirt, shorts, and a wig to channel Roger’s forehand.

The new framework, AdaPT, uses speed-adaptive planning to bridge sim-to-real and replicate elite kinetic signatures.

[Humanoids daily @humanoidsdaily]

15 feet up, 15 feet down.

Figure has demonstrated the Figure 03 autonomously climbing and descending a 15-foot industrial ladder. CEO Brett Adcock confirms the vertical locomotion milestone isn't just a lab stunt—it’s mapped directly to an active mezzanine customer use case.

Full breakdown:
https://t.co/f6TIbAfxmw

[Humanoids daily @humanoidsdaily]

https://t.co/TNqz7Fio7p

[Tao Huang @TaouHuang]

🎾 What if humanoid robots could play tennis like @rogerfederer, @RafaelNadal, @DjokerNole, and others?

🤖Introducing our project AdaPT: Adaptive Motion Planning and Tracking, which made this possible!

Website: https://t.co/mLkEMZSeWb
Youtube: https://t.co/8nSyeArxho

[Humanoids daily @humanoidsdaily]

Researchers taught a Unitree G1 humanoid to play tennis using broadcast footage of Federer, Nadal, and Djokovic—and then naturally put it in a polo shirt, shorts, and a wig to channel Roger’s forehand.

The new framework, AdaPT, uses speed-adaptive planning to bridge sim-to-real and replicate elite kinetic signatures.

[Humanoids daily @humanoidsdaily]

Full breakdown on the funding round, hardware specs, and XPENG’s production roadmap:
https://t.co/1WdzZ8idn5

[Humanoids daily @humanoidsdaily]

XPENG’s robotics division has closed a $900M+ funding round at a post-money valuation exceeding $6.3B, marking the largest single-round private raise in China’s physical AI sector.

Key details:
• Led by IDG Capital with Gaorong Ventures, plus strategic backing from Alibaba & Tencent

• XPENG retains majority control

• Latest IRON specs: 76 DoF body, 21-DoF hands, 2,250 TOPS on-device compute (3x Turing AI chips)

• Mass production slated for late 2026, targeting initial deployment in XPENG showrooms before global delivery in 2027

[Humanoids daily @humanoidsdaily]

Wang Xingxing called it.

Sub-10s was shattered at the World Robot Games with a 9.32s 100m sprint by a robot from TianGong Robotics.

[Humanoids daily @humanoidsdaily]

Will a humanoid robot outrun Usain Bolt by this summer? 🏃‍♂️🤖

At the 2026 Yabuli Forum, Unitree CEO Wang Xingxing predicted sub-10s 100m sprints by mid-year.

[Humanoids daily @humanoidsdaily]

They out-sprint us and out-jump us, but Messi and Ronaldo don't need to lose any sleep just yet. It'll take at least another year.

[Humanoids daily @humanoidsdaily]

9.32 seconds 💨 Beating Usain Bolt's 100m world record (9.58)

[Humanoids daily @humanoidsdaily]

Booster Robotics at the World Humanoid Robot Games Opening Ceremony

[Humanoids daily @humanoidsdaily]

@adcock_brett Impressive! Is this capability mapped to an active commercial pilot requirement, or is the team treating this as a general foundational mobility milestone for the fleet?

[Brett Adcock @adcock_brett]

15 feet up. 15 feet down. Fully autonomous

[Humanoids daily @humanoidsdaily]

@Kenny___Rose @tomsanocki @1x_tech Looking forward to getting more details.

[Humanoids daily @humanoidsdaily]

Generalist AI released GEN-1.5 yesterday — a robot foundation model that can pick up a new task from a single demonstration of a few seconds, with no training at all.

Here it is asked to put a block in a bowl. Someone has covered the bowl with a piece of paper.

So it removes the paper, drops the block in — and then sometimes puts the paper back on top.

Why that matters: a robot taught the normal way is replaying a motion it was shown. Put an obstacle in the way and it just runs the motion into the obstacle. It has no notion that the paper is a problem, because the paper wasn't in the demo.

GEN-1.5 was never shown the paper either. It worked out on its own that the bowl needed clearing first: