Robotics

Why the hardest part of physical AI isn’t seeing the world — it’s closing the loop back to action.

Post was made on 2nd Jan 2026 Who is Brett Adcock and Why Does This Prediction Matter? Brett Adcock isn’t just another tech founder making bold claims on LinkedIn. He’s a serial entrepreneur who previously founded Archer Aviation (an electric air taxi company now publicly traded) and Vettery (a hiring marketplace acquired for $100M+). In 2022, he founded Figure AI with a singular mission: build the world’s first commercially viable humanoid robot. ...

The Intelligence Delusion I think We’ve been pursuing only half of intelligence for seventy years. When Alan Turing proposed his famous test in 1950, he made a brilliant and practical choice. Faced with the impossible question of “what is consciousness?”, he focused on something measurable: could a machine engage in conversation sophisticated enough to fool a human judge? It was an elegant solution to a philosophical problem that had no clear answer. ...

Reinforcement learning (RL) has led to impressive robot demos, but deploying these controllers in the real world remains challenging. Training a policy in simulation is one thing, making it work reliably on a physical robot is another. The hidden costs include months of fine-tuning, poor generalization across tasks, and an unpredictable trial-and-error process. Credit: π0: First Generalist Policy by Physical Intelligence (π) A recent paper on π0: A Vision-Language-Action Flow Model for General Robot Control introduces a foundation model for robotics, designed to overcome these limitations. Instead of training separate RL policies for each task, π0 is pre-trained on over 10,000 hours of diverse robot data across multiple platforms and 68 tasks. It builds on pre-trained vision-language models (VLMs) to incorporate internet-scale knowledge, allowing it to follow language instructions and generalize more effectively to new scenarios. ...

AI has made extraordinary progress in multimodal learning, particularly in aligning text and images into shared embedding spaces. Models like CLIP and GPT-4V(ision) are prime examples of this success, showing how architectures can bridge the gap between vision and language. But when we add videos to the mix, the landscape changes dramatically. Despite existing architectures and innovations, aligning text, images, and videos into a common embedding space is still imperfect and incomplete. ...