PRESENTED BY
Render labeled “gen3,” extracted from Tesla’s Android app by Humanoids Daily. It does not confirm the final hardware design.
A possible first look at Optimus Gen 3 came from an unexpected place: Tesla’s own app. After the renders began circulating, we checked the Android files ourselves and confirmed images labeled “gen3.” They offer clues to the robot’s next design, though they don’t confirm its final appearance. Our follow-up also found that the images had been in the app since August 27—weeks before they attracted widespread attention.
This week also brought a reported jump in Optimus production, AMD’s $8.2 billion World Labs deal, and 140 years of simulated robot soccer. We spoke with the founders of Feather, O-ID and Delta Intelligence about what they are actually building—and why. Stay to the end for Figure 02’s possible Terminator-style farewell.
IN THIS EDITION
Tesla: what our app investigation found, and the manufacturing challenges behind rising output.
AMD and Skild: an $8.2 billion spatial-AI deal and 140 years on the soccer pitch.
Our reporting: Feather’s developer platform, O-ID’s repairable robot and Delta’s industrial ambitions.
New arrivals: Vesoma steps out of stealth; Sharpa brings three products to IROS.
And finally: Arnold Schwarzenegger has a retirement plan for Figure 02.
TOP STORIES
We weren’t first to report the images. Our contribution was to verify the files and trace their history: a comparison of four Android builds dates the three “_gen3” assets to August 27. Higher-resolution versions arrived September 5 and remained byte-identical in the September 22 release.
The images show a more enclosed exterior around the hips and lower legs. But they live in an app directory labeled “mock”: they establish what Tesla packaged in its software, without confirming the final robot design, a usable Optimus interface or a launch date.
The Information reports that Tesla built several hundred robots a week in August, up from a few dozen in the second quarter. Most reportedly support internal testing, training and data collection. The V3 machines now being built still differ from the intended customer version.
The same report describes hand-assembly problems, touch-sensor failures and inconsistent supplier quality. Rising output is a meaningful development; making those units durable and useful enough for customers remains the harder test.
SPONSORED BY MATRIX
Giving Physical Intelligence a Sense of Touch
Better robot learning starts with better human data. Matrix’s motion-capture system pairs a soft sensor glove with a depth camera to capture hand movements—even when parts of the hand move out of view. Explore the Matrix data glove range for your next data-collection setup.
AMD has agreed to acquire Fei-Fei Li’s World Labs in an all-stock deal valued at approximately $8.2 billion. Closing is expected by the end of 2026, subject to approvals. Li would become AMD’s executive vice president and chief scientist.
For robotics, the attraction is spatial intelligence: models that reconstruct and generate 3D environments for simulation and training. The deal would bring that research inside a chipmaker, creating a closer connection between the models and the hardware running them.
Skild AI’s soccer experiment explores learning through simulated competition. Image: Skild AI.
Skild says S1 accumulated more than 140 years of simulated self-play, competing against recent versions of itself in NVIDIA Isaac Sim. The company reports that the resulting policy can play against humans and other robots on physical hardware, with early passing and coordination in four-agent games.
The training sequence matters: human-referenced dribbling and kicking drills came first; goal-scoring then drove self-play without additional demonstrations. The wider promise is learning through experience without needing a fresh human example for every tactic. Whether that improves commercial work is still an open question.
OUR REPORTING · THREE FOUNDER CONVERSATIONS
Feather’s developer platform starts at a promotional $29,990. Image: Feather.
Feather launched with $7.6 million in funding and a wheeled humanoid starting at a promotional $29,990. It supplies the robot, controls and software development kit; customers choose the AI models and applications.
In our interview, co-founder and CEO Hoa Mai argued for an “iPod moment” before an “iPhone moment”: keep the machine simple enough to do useful work with today’s AI. He says manufacturing is underway in the Bay Area, with current capacity for 30 robots a month. The proposition puts affordable, dependable hardware at the center—and leaves developers room to build.
Tokyo-based O-ID raised $1.2 million for a modular industrial robot with replaceable joints, limbs and compute. In written answers to us, CEO Stian Jakobsen explained the planned repair workflow: power down, swap the worn module, load its stored calibration and run readiness checks.
The goal is maintenance that factory staff can handle without waiting for a visiting engineer. Predictive wear monitoring and the promised repair times remain development targets. Our follow-up gets into the service model behind the modularity pitch, and why the founders chose Japan.
Delta’s Δ₀ demonstration includes operating a turntable. Screenshot: Delta Intelligence.
Delta Intelligence’s Δ₀ coordinates 69 degrees of freedom on a modified Unitree G1, with demonstrations including bed-making and opening a dishwasher. Its report describes dishwasher success rising from 23% to 89% as training-data usage increases.
But founder Xiaojian Ma told us the near-term targets are energy and infrastructure: underground facilities, offshore sites and other difficult workplaces. Our written interview explains the planned head-and-backpack system for other humanoid platforms, why whole-body data collection was part of the strategy from day one, and what comes next. The research results are promising; industrial reliability is the next test.
PRESENTED BY ROBOSTRATEGY
Most of the companies we usually cover in this newsletter are still private, which puts them out of reach for public-market investors. RoboStrategy, Inc. ($BOT) is built to give you access to several of them.
Investing involves risk, including possible loss of principal. Read the prospectus before investing.
ALSO THIS WEEK
The European team behind the Kyle walking prototype has introduced Vesoma, with more than 60 people across Munich and Limassol. Former Google DeepMind research director Martin Riedmiller is chief AI officer, alongside founders Nikolai Ensslen and Peter Skoromnyi.
Its bet is that robots should improve through their own physical interactions. The first Vesoma 1 units are being built; the polished launch images are explicitly design visualizations, with manipulation performance and customer delivery timing still to come.
Sharpa’s new D01 brings dual arms and body tactile sensing; the W02 hand weighs under 750 grams with 21 active degrees of freedom; and the AE01 glove captures hand movement and returns fingertip vibration feedback. Together, they span the robot, its sense of touch and the operator’s interface. Public prices and delivery dates were not specified.
AND FINALLY…
Figure needed room. An internet poll favored blowing the retired robots up. Then Arnold Schwarzenegger suggested melting them—and Brett Adcock sent a DM.
Now an AI-generated farewell teaser points to September 30. It suggests a Terminator 2-style sendoff, rather than showing an actual melting or confirming Arnold’s attendance. Apparently, disposal advice from the Terminator carries some weight.
This edition of Humanoids Daily was presented by RoboStrategy.
Thanks also to Matrix for supporting this week’s edition.
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.