September 1, 2026Introducing Atlas, our new omni world model for spatial intelligence. Atlas: A World Model for Spatial Intelligence World models generate, reconstruct, and simulate any possible world. They understand how worlds appear, behave, and evolve so that we can render imagined worlds for creative users, simulate the real world in high fidelity, and help robots plan actions. At World Labs, we build these general purpose world models in pursuit of spatial intelligence. Today we are introducing Atlas, our next-generation world model. Atlas is an omni model that we pretrained from…

A 36B parameter model just blurred one of robotics’ biggest boundaries, the gap between seeing a scene and acting on it. Perceptron AI has released Isaac 0.5, an open-weight Embodied AI model designed to use one backbone to understand video, reason about a physical task, and generate robot actions. One model connecting perception, reasoning, and robot control. @perceptroninc

$1 million in sales in under SEVEN hours, is the Microduck really that impressive? Here are other affordable open-source robots for context (+ my thoughts): ▸ @LeggedRobot Mini Pupper 2 (~$649): ROS2, SLAM/navigation, quadruped ▸ @ROBOTIS MINI (~$500): 16-DoF humanoid, smart servos, full-body motions ▸ @PetoiCamp Bittle X (~$319): cheapest option, self-balancing quadruped, beginner-friendly ⭐️@pollenrobotics + @huggingface releases: ▸ Reachy Mini ($399–499): vision + voice interaction, expressive head, AI assistant focus ▸ Microduck ($399): learned biped walking, onboard AI policies (out of…

Humanoid robots just took tennis to another level. 🎾🤖 Galbot robots completed 100+ consecutive rallies at the World Humanoid Robot Games 2026 while tracking the ball, moving around the court and responding in real time. This is not just about hitting a tennis ball. It combines vision, prediction, balance, whole-body control and real-time decision making. Would you play against a humanoid robot? 👀 🎥 Media: @GalbotRobotics ⚠️ This content is shared for informational purposes only. CTO Robotics Media is a media platform and does not own or develop the technology shown. Credit belongs to the…

3:01 PM ETFigure

Figure founder Brett Adcock has posted video of the company's humanoid climbing 15 feet up a ladder and 15 feet back down with no human input, three weeks after its first autonomous ladder climb. Figure credits its Helix AI system, whose System 0 layer runs balance and contact control a thousand times a second using stereo camera vision alongside the robot's own sense of body position. Figure has not released the technical details and the demonstration has not been independently verified.

4:12 PM ETGemini Robotics

Announced July 30, 2026 (author Carolina Parada) as part of Gemini Robotics 2, three models: Gemini Robotics 2 (VLA) converts vision+language into motor control first DeepMind VLA to control a full humanoid feet-to-fingertips under one checkpoint. Partner-only. Gemini Robotics ER 2 (VLM) the embodied-reasoning brain; publicly available. Gemini Robotics On-Device 2 (VLA) runs locally, adapts to new embodiments in a few hours. Partner-only.

6:10 AM ETGoogle

I’ve said before that I think Google & DeepMind are playing a longer game with models. This is still a firm stance. Because a lot of the time, i think people look at Google’s model strategy too narrowly; Every Gemini release immediately gets reduced to how it compares with other labs or whatever benchmark is trending that week. Meanwhile the it’s more about where all of that intelligence might show up next. Like in Reasoning, Vision, Video, Agents. And particularly robotics. I believe Gemini Robotics ER 2 is a clear example here. It was built on Gemini 3.5 Flash. Except, now its intelligence…

Humanoid BOM model: Current humanoid BOM = ~$68k based on channel checks across 3 humanoid OEMs (UK and US). - Actuation = 33.8% of current BOM - Dextrous hands = 30.3% - Structure = 7.6% - Sensing = 7.5% - Electrical = 6.0% - Vision = 4.7% - Compute = 4.6% - Power = 4.6% - Other = 0.9% * Piecing together various component costs across different Western OEMs. Numbers may be over/understated for this reason.

The Unitree G1 humanoid robot can drive a go-kart, controlling the steering wheel and pedals around an indoor track. It can handle tight corners and even perform controlled drifts. Developed with Symbiosis Robotics Direct Perception Control system, the G1 uses vision, body state, and action history to coordinate its movements in real time.

Instead of choosing between training a world model and training a language conditioned robot policy, why not do both? LDA-1B s a new foundation model that is trained on 30,000 hours of human and robot interaction data. Part of the secret is that LDA-1B jointly learns forward dynamics, action prediction, and visual forecasting, all in a structured DINO latent space which avoids the pitfalls of redundant pixel-level prediction which isn’t necessarily aligned robot action. This approach works on both dexterous hands and simple robot grippers; it also generalizes across objects, tasks, and…