All articles
Design Arena raises $7.9M for human taste in AI models
Design Arena raised $7.9M in a seed round led by Index Ventures. The platform collects human judgments on AI aesthetics — 5.3M users and $60M ARR, with frontier labs as customers.
Musk spends half of Tesla's earnings calls on AI and robots
Elon Musk now spends about half of Tesla's earnings-call time on AI, robotaxis and humanoids, according to a Hudson Labs analysis. In 2022 those topics took up 15–20 percent of the conversation.
Walden Robotics teams with Toyota on practical humanoids
Walden Robotics — a spinout from Toyota Research Institute — emerged from stealth with $300M at a $1.1B valuation. It builds wheeled humanoids for factories, with Toyota as investor, partner and first customer.
T-Rex: a manipulation model that reacts to touch
A team from Stanford, UC Berkeley and NVIDIA — with Li Fei-Fei among 34 authors — unveiled T-Rex, a vision-language-action model that adds high-frequency touch. It lifts success rates by over 30 points across 12 dexterous tasks.
NVIDIA Cosmos 3: one world model for physical AI
At SIGGRAPH 2026 NVIDIA showed Cosmos 3 — a physical-AI foundation model fusing understanding, world generation and action. The edge Cosmos 3H (4B params) runs on Jetson, and Cosmos Dreams cuts simulation cost from 64 GPUs to one.
DEEP Robotics: a quadruped robot with AI for warehouse management
China’s DEEP Robotics showed its Lynx S10 quadruped paired with an AI foundation model: it inventories the warehouse, scans QR and RFID tags, and patrols the floor for fire and accidents.
Reimagine Robotics: robots that learn on the factory floor
A startup founded by four ex-Google DeepMind engineers left stealth. Its robots learn tasks directly from workers, no programmers required — cutting prototype build time from a day to 10 minutes.
Zhiyuan (AgiBot) preps a Hong Kong IPO and enters the factory phase
Zhiyuan (AgiBot) is targeting a Hong Kong listing at a 363–433 billion yuan valuation. It is deploying humanoids on an electronics line and building its own robot-control models.
MiniMax H3: an open video model aiming to be multimedia's DeepSeek
MiniMax unveiled H3, a multimodal generation model unifying text, image, video and audio, with open weights promised. The firm is aiming to become the "DeepSeek" of video generation.
Unitree warns US curbs could cap its growth ahead of IPO
Ahead of its Shanghai STAR Market debut, Unitree warned in its filing that new U.S. curbs on advanced foreign robots could cap its growth. International markets make up over 40% of sales, with the U.S. alone at 13–19.5% of revenue.
VLASH: MIT's method that lets robots think while they move
MIT researchers unveiled VLASH, a method that lets robots plan their next move while still executing the current one. The result is up to 11.8x lower reaction latency for VLA models, with no architecture changes and no extra compute.
Gemini Robotics ER 2: a high-level brain for robots
Google DeepMind released Gemini Robotics ER 2 — a reasoning model that plans multi-step tasks, analyzes video in real time and coordinates multiple robots, delegating motion to separate VLA models.