How AI breakthroughs like diffusion models, visual language models, and liquid neural networks are transforming the way robots learn to move and pick up skills
Advances in physical AI mean machines are learning skills previously thought impossible — “This is not science fiction,” … X: @alisonkilling , @samjoiner , and @leokelion X: @alisonkilling : NEW: Buttering toast, picking up a can, tying shoelaces - tasks that are easy for humans, but fiendishly difficult for robots. But now AI is allowing them to learn this. Are the robots finally coming? https://ig.ft.com/... [video] Sam Joiner / @samjoiner : NEW: Are the robots finally coming? In our latest visual story, @upyorkshire @sam_learner @inari_ta @ian_bott_artist @Mikepeeljourno @peterjandringa @JALWilliams_ @carolinenevitt @Dan_Clark5 explain how AI is powering a robotics revolution. 🤖👉 https://ft.com/... [video] Leo Kelion / @leokelion : What started with chatbots is coming to robots.... This stonkingly brilliant robotics visual explainer in today's online FT conveys how large language models are being applied to the physical world. Grab a coffee and enjoy https://ig.ft.com/...
Context & Ripple Effects
Recent coverage has focused on moving robotics from hand-coded behavior toward models trained on physical-task data: Physical Intelligence's approach centers on feeding robot task data into an AI model, while DeepMind's AutoRT work applied visual-language models to situational awareness.
This report places diffusion models, visual-language models and liquid neural networks in that same arc, with the emphasis shifting to movement and everyday object manipulation rather than language understanding alone.
First-order effects
- Robot developers can use newer AI approaches to train movement, perception and manipulation skills that previously required more bespoke programming or were difficult to automate reliably.
- The immediate competitive focus moves toward systems that can translate visual understanding and learned physical behavior into practical task execution.
Second-order effects
- Robotics companies pursuing general-purpose models, including Covariant's physics-processing RFM-1, face pressure to show that their models improve real-world manipulation rather than only narrow task performance.
- Demand shifts toward the data, evaluation environments and deployment workflows needed to teach and validate physical skills, not just model development.
Third-order effects
- If these methods generalize across tasks, robotics competition could increasingly be organized around reusable physical-world models and the task data that improves them, rather than one-off automation scripts.
- The limiting question becomes whether learned behavior is dependable enough in varied operating conditions; that will determine how quickly physical AI complements conventional automation.
The trend: Robotics is moving from specialized, explicitly programmed machines toward physical AI systems that learn transferable perception-and-manipulation skills from data.