Servoservobrief.com, Tuesday, November 3, 2026
A wave of Chinese embodied-AI money landed, led by training-data and humanoid raises, while Google and Meta's LeCun pushed world models on the research side.
-
GenRobot AI raises $100M to sell robot training data at scale
The Beijing startup, founded in 2025 by an ex-Momenta director, disclosed a $100M Series A+ that closed in Q3, with no lead or valuation named. It sells wearable rigs that record a worker's vision, hand and body motion, and claims over one million hours of human demonstration data across 30-plus AI customers (vendor-reported). Humanoid maker Galbot joining the cap table shows robot builders paying to lock up the data their own models train on.
Read more on Seedtable / TechNode -
ROBOTERA closes over $200M for humanoids
The Chinese humanoid maker raised more than $200M in a round led by SF Group. No valuation or detailed use of proceeds disclosed, and the figure is vendor-reported. It adds to a run of large China humanoid raises, keeping Chinese hardware developers near the top of the sector's capital inflows.
Read more on Funding wire -
RoboScience raises several hundred million yuan in Series A2
The Chinese embodied-AI startup closed a Series A2 worth several hundred million yuan (roughly tens of millions of dollars), co-led by Hupo Capital, Hongruida and Shanghai Semiconductor Industry Investment. It is the company's second large raise this year after a similarly sized pre-A in February, a pace that signals heavy strategic and industrial appetite for Chinese robotics platforms.
Read more on Dealroom -
Google ships Gemini 4 Argon, a model tuned for robots
DeepMind released Gemini 4 Argon for Robotics alongside its new Gemini 4 frontier line, a model variant aimed at powering robot perception and control. No benchmark numbers or robot partners were detailed in the listing. It extends Google's push to make general foundation models the brain for physical machines, putting it against Physical Intelligence and Nvidia for the robotics model layer.
Read more on Google DeepMind -
LeCun's lab proposes hierarchical world models for planning
A paper from Yann LeCun, Lerrel Pinto and collaborators introduces hierarchical world models with implicit dynamics, a way for robots to predict outcomes at multiple time scales to plan actions. A world model is a learned simulator of how the environment responds. The work pushes LeCun's bet that prediction-based models, not language models, are the path to capable embodied agents.
Read more on arXiv cs.RO