Gemini 4 puts Apollo on target — and PI opens its API

PRESENTED BY

Watch a person throw a beanbag. Build a simulation. Practice, adjust, then try it on a real humanoid. Google’s Gemini 4 Argon demonstration takes Apollo through that process—and, Google says, finishes with seven consecutive cornhole shots.

It’s a vivid example of a broader shift in how robot skills are developed. Physical Intelligence is opening an API for model training and inference, while manufacturers are working on the less glamorous requirements: endurance, charging and factory integration. Unitree’s rally—and its own warning about commercialization—capture the gap between expectations and the work still ahead.

IN THIS EDITION

LEAD STORY

Google says Apollo 2 sank seven consecutive cornhole shots after Argon used human video, MuJoCo simulation and feedback from physical attempts to develop the movement. The workflow includes mapping the arm swing, adjusting the stance and refining the release when bags miss.

The interesting part is how those steps connect. A general-purpose model helps construct the practice environment, inspect failures and revise the action. Google says this can shorten weeks of manual simulation modeling to hours.

The seven-shot streak is a demonstration, not a disclosed overall success rate: total attempts and the human engineering involved weren’t detailed. And Argon’s reported 87% ERQA score measures image-and-text reasoning questions, separately from performance on a physical robot. Our breakdown separates the real-hardware results from simulation and generated-video examples.

See the throw—and how the workflow works →

PI’s new API beta lets teams upload demonstrations, fine-tune models on PI-managed infrastructure and use hosted inference. The pitch is practical: spend less time maintaining training infrastructure and more time testing a useful task.

One example is Almond’s Axol robot deburring parts for its own assembly, which PI says took four days of API iteration to reach. Teams still need to collect data, evaluate on hardware and address failures. Most showcased applications used cloud inference, with connectivity challenging some teams. Access is application-based, with users onboarded in batches.

What the API offers developers →

HARDWARE & DEPLOYMENT

Boston Dynamics’ new prototype-shop film shows engineers and machinists working side by side, revising precision parts and feeding lessons back into the next build. One employee describes roughly a day to build a robot today, with an hour as next year’s ambition. That is a build-time claim and target—not a disclosed shipment rate.

Go inside the Atlas workshop →

Brett Adcock says a Figure 03 has walked on a treadmill for nine straight days, covering about 560 kilometers. The test began as a side project to expose failures and was still running when he posted the update.

That is a striking amount of repeated motion. It is also specifically a walking-endurance result: the power arrangement, maintenance and intervention log weren’t disclosed, so it should not be read as nine days on one battery or a measure of reliability across factory tasks.

Watch Figure’s endurance test →

KAIST’s Humanoid v0.9 demo combines balance, body movement and arm forces to push objects up to 35 kg and lift up to 20 kg. The DRCD Lab and arm-development partner RAILAB combine onboard visual feedback with skills learned through reinforcement learning and human demonstrations. These are two different handling tasks, rather than a single payload rating.

Watch KAIST’s whole-body demonstration →

At its new Taicang facility, Schaeffler says six humanoids work across five stations, preparing, assembling, inspecting and packing components. The roughly 7,500-square-meter site combines manufacturing with training, data collection and digital twins.

“Humanoids making humanoids” is the headline, but the announced output is joint modules and other key components. Schaeffler also reports around €400 million in incoming humanoid-robotics orders, rather than recognized revenue. Throughput and intervention rates for the new line haven’t been disclosed.

Inside the Taicang facility →

AGILITY WATCH

Agility has packed several updates into the run-up to its planned Nasdaq listing. Here’s what changed since last week:

MARKETS & BUSINESS

Unitree closed up 16.15% on October 19, alongside a broader robotics rally and new Chinese policy support for robot applications. Its latest trading disclosure adds a caution: technology maturity, customer demand and supply-chain capacity could constrain commercialization.

The company put its trailing price-to-earnings ratio at 388.25, versus a broad equipment-manufacturing industry average of 39.51. It reported normal operations and no undisclosed material information. This is a valuation and commercialization warning, rather than an announced operational setback. Unitree’s October 22 conference is the next event to watch.

Read Unitree’s warning →

Musk says Optimus and Cybercab will represent the vast majority of Tesla’s value, without giving a valuation or timeline. Separately, Tesla’s board is seeking the re-election of Musk, Robyn Denholm and JB Straubel at the December 3 shareholder meeting. Its letter points to work on Fremont’s Optimus line, but adds no output figures or delivery date.

Musk’s outlook and the Texas factory update →
What the shareholder letter says →

Swedish startup 120ft proposes three 40-foot containers: one for preparing components, one for assembly, and one for inspection and packing. Brands would pay per finished product, with production runs pitched from 100 units.

Early pilots focus on rigid parts under 5 kg, assembled with screws or press-fits. Its angel funding supports a move toward commercial production. The appeal is smaller batches without owning a factory; the test is delivering them reliably. The company’s seven-minute skill-programming ambition is a 2027 target, not today’s production capability.

Explore 120ft’s microfactory model →

OUR REPORTING

We put questions to the team after the announcement of an approximately $200 million all-stock transfer of FFAI’s robotics business to FFR, an existing listed company. It is not a $200 million cash injection; FFAI expects to retain about 65.7% ownership after the transaction and financing.

The team confirmed that partner hardware remains central and its U.S. production roadmap is unchanged. But its 817 devices “sold and/or shipped” figure does not distinguish completed customer deliveries from other categories, and it supplied no revenue total through September. Our Q&A examines what the answers clarify and what remains unspecified.

Read our Faraday Future follow-up →

FOR BUILDERS

RoboParty’s RP1 release is now more than a roadmap. Public repositories span mechanical designs, manufacturing files, electronics, robot models and development tools. MimicLite lists motion-policy checkpoints for two RP1 configurations. Scope and licensing vary by component; our guide explains what developers can actually use.

AND FINALLY…

Dynamic Creatures’ furry, wide-eyed character combines a 43-actuator platform with animatronic design. The office demo is charming, and the company says it is developing a Vision-Language-Expression model to connect what Hugo sees and hears with gestures and movement.

How much of the clip is autonomous remains unclear, and navigation and safety for close interaction are still in development. For now, it’s a memorable glimpse of a robot designed to make an encounter feel personal.

Watch Hugo say hello →

This edition of Humanoids Daily was presented by RoboStrategy.

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.