RoboBrain 2.0

Dear Readers,

How long have we been talking about “intelligent robots” – and then settling for talking vacuum cleaners and stiff gripper arms? This week brings a breakthrough of a different kind. With RoboBrain 2.0, a new chapter in robotics is beginning: one in which machines not only execute commands, but can understand their environment, plan sequences of actions, integrate feedback, and react flexibly – visually, spatially, and linguistically.

What particularly impresses me is that RoboBrain is not only powerful, but also radically open. Instead of hiding behind black boxes, the team relies on open source – including the 7B model, inference code, and soon a 32B release on Hugging Face. This not only builds trust, but also democratizes access to robotic intelligence for developers, universities, and startups worldwide.

In Today’s Issue:

All the best,

RoboBrain 2.0 is an open-source multi-modal large language model (MLLM) that equips robots with advanced spatial perception and planning. It analyzes multiple images or videos, understands complicated instructions such as “grab the cup,” plans long sequences of actions, and adapts flexibly during execution—with closed-loop feedback and structured memory. This is a game changer for the AI community. Behind the scenes, it works with benchmarks such as BLINK-Spatial and RefSpatial, even outperforming large closed-source models with over 50% better performance in visual spatial reference!

What makes it special is that it is completely open source – including a 7 trillion parameter model on Hugging Face, simple inference code, and soon even a 32 B model. So anyone experimenting with robots gets a powerful toolkit that combines visual understanding and action competence like never before.

Outlook: Imagine giving a household robot a plan – and it ticks, talks to you, grabs things, understands its environment – all in one model. How could that change the way we work and live?

Why it matters: This project combines visual intelligence with action competence in an open-source system. It accelerates research and makes advanced robot AI accessible to everyone.

Sources:

🔗 https://github.com/FlagOpen/RoboBrain2.0

🔗 https://superrobobrain.github.io/

In 2 years you will be working for AI

Or an AI will be working for you

Here's how you can future-proof yourself:

Join 1,000,000+ pros at companies like Google, Meta, and Amazon that are using AI to get ahead.

Sign up and start learning AI

In a world-first, researchers from Insilico Medicine and the University of Toronto used a hybrid quantum-classical AI system to design and experimentally validate new drug leads for the notoriously "undruggable" KRAS cancer protein. The team leveraged a quantum computer to generate novel molecular structures, which were then filtered and ranked by a classical AI before being synthesized and tested in a lab. While true quantum advantage has not yet been demonstrated, this success provides a foundational template for integrating quantum hardware into real-world drug discovery pipelines to tackle previously intractable diseases.

“Benchmark comparison across spatial reasoning and task planning. RoboBrain2.0-32B achieves state-of-the-art performance on four key embodied intelligence benchmarks: BLINK-Spatial, CV-Bench, EmbSpatial, and RefSpatial. It not only outperforms leading open-source models such as o4-mini and Qwen2.5-VL, but also surpasses closed-source models like Gemini 2.5 Pro and Claude Sonnet 4 — especially in the challenging RefSpatial benchmark, where RoboBrain2.0 shows a >50% absolute improvement.”

According to Nvidia CEO Jensen Huang and OpenAI CEO Sam Altman, humanoid robots could go into series production as early as 2027: Advances in AI models and simulations are enabling autonomous, physically capable machines – from logistics to manufacturing.

Nvidia is developing special foundation models for robotics. Altman even predicts self-assembling robot fleets. The potential: faster time to market, profound transformation in industrial and service processes, and a foundation for disruptive automation.

Toyota is opening its ”Woven City" at Mount Fuji in the fall: a 175-hectare laboratory for AI, robotics, autonomous driving, and smart homes. One hundred pioneers will spend ten months living in real-life conditions, including emission-free mobility, robot assistance, and health monitoring via smart homes. Relevance: Real-world testing of urban technology ecosystems with global appeal and innovative momentum.

Amazon is testing humanoid delivery robots in a realistic “Humanoid Park” in San Francisco – including an obstacle course made up of kilometer markers. The goal: to deliver packages autonomously from Rivian vans and relieve drivers.

AI software is still crucial, with hardware sourced externally. Obstacles remain: variable environments with pets and children. Nevertheless, this signals a massive step toward autonomous logistics in urban areas.