Short Wave

AI is great at predicting text. Can it guide robots?

February 11, 2026

Key Takeaways Copied to clipboard!

  • While AI chatbots benefit from vast internet data, applying this intelligence to robotics faces significant hurdles due to the lack of comparable real-world training data and the complexity of physical tasks. 
  • Researchers are exploring simulation as a way to rapidly generate training data for robots, but perfect simulation of real-world physics and object interaction remains a major challenge. 
  • AI is currently finding success in robotics by being applied to specific, isolated tasks like image recognition for package sorting, rather than achieving general-purpose intelligence for complex actions like making a sundae. 

Segments

AI Moving From Virtual to Reality
Copied to clipboard!
(00:00:21)
  • Key Takeaway: AI’s increasing presence in virtual life is now transitioning into physical applications, exemplified by major tech companies showcasing AI-powered humanoid robots like Tesla’s Optimus and Google’s Gemini-powered agents.
  • Summary: Artificial intelligence is rapidly expanding beyond digital interfaces into the physical world. Companies like Tesla and Google are actively developing humanoid robots powered by advanced AI software. This shift raises questions about the practical capabilities of these physical agents compared to their virtual counterparts.
Stanford Lab and Teachable Robots
Copied to clipboard!
(00:03:51)
  • Key Takeaway: Robots at Stanford’s IRIS Laboratory use AI models like Open VLA, allowing them to be taught tasks through repeated demonstration rather than explicit, detailed programming.
  • Summary: The IRIS Laboratory at Stanford focuses on AI-powered robotics, moving away from traditional, hard-coded instructions. Graduate student Moojin Kim demonstrated a robot arm powered by Open VLA, a teachable neural network. Tasks are learned by reinforcing connections within the network through repeated physical demonstrations, similar to how large language models learn.
The Dream of General Robotic Intelligence
Copied to clipboard!
(00:06:27)
  • Key Takeaway: The long-term goal for AI robotics researchers, like Chelsea Finn, is developing software enabling robots to intelligently execute simple, natural language commands for everyday tasks like making a sandwich or cleaning a kitchen.
  • Summary: Chelsea Finn, who runs the Stanford lab and co-founded Physical Intelligence, aims for robots that can operate intelligently in any situation based on simple commands. This includes performing basic household or service tasks such as folding laundry or restocking shelves. Physical Intelligence recently demonstrated a mobile robot successfully folding laundry taken from a dryer.
Hurdles: Data Scarcity and Task Complexity
Copied to clipboard!
(00:07:55)
  • Key Takeaway: The primary barrier to widespread AI robotics success is the massive data requirement for physical tasks, which is orders of magnitude greater than for text prediction, leading to robots getting stuck or making errors in real-world execution.
  • Summary: Despite successful demonstrations, real-world robotic tasks often cause confusion, mistakes, and getting stuck, requiring human cleanup. Unlike chatbots trained on the entire internet, robotics lacks foundational online data for training, meaning current manual teaching rates would take 100,000 years to accumulate sufficient data. Furthermore, physical tasks involve complex forces and object interactions that are difficult to perfectly simulate.
Framing the Problem Differently
Copied to clipboard!
(00:11:09)
  • Key Takeaway: Some experts argue that the fundamental problem framing for robotics—predicting complex physical actions—is inherently more difficult than the next-word prediction task central to successful AI chatbots.
  • Summary: Matthew Johnson Roberson suggests the issue isn’t just data quantity but the framing of the problem itself. Predicting the next word in a sentence is a relatively simple task compared to the multifaceted requirements of robot movement and interaction. Current data collection methods may be insufficient unless researchers find a fundamentally different way to teach robots complex physical skills.
Near-Term Practical AI Integration
Copied to clipboard!
(00:12:42)
  • Key Takeaway: Near-term progress in AI robotics will likely involve integrating AI for specific sub-problems, such as vision or grasping, rather than achieving immediate, full-scale general intelligence.
  • Summary: AI is already proving effective in specific robotic components; Ken Goldberg’s package sorting company uses AI image recognition to optimize package grasping points. This targeted application is showing real success in industrial settings. Progress is expected to arrive incrementally, solving parts of the robotic challenge before a complete, universal solution emerges.