Industry Social Media and Updates

Video‑trained Robots Achieve Ten‑fold Faster Control

What if robots learned from watching videos instead of being programmed? That's the bet Rhoda AI just made with $450 million. Most robot companies teach machines "what things are." They use vision-language models trained on internet images and text, then fine-tune them on robot demonstrations. It works. But there's a latency problem—the robot's decision loop runs at 1-2 cycles per second. When timing matters (welding, precision handling), that's slow. Rhoda's doing something different. They pre-train on internet video to teach robots how the physical world actually moves. Then they predict what the next few frames will look like and work backwards to figure out what actions produce those frames. The result: 10 Hz closed-loop control instead of 1-2 Hz. Same foundation model idea. Different architecture. Different speed. The team matters too. Jagdeep Singh built QuantumScape (now public) and founded Infinera (sold to Nokia for 2.3B). Gordon Wetzstein from Stanford. Eric Chan from WorldLabs on generative models. These aren't people experimenting. They're executing. The Real Test Does it actually work better in real factories with real variability? We won't know for 12-18 months. But when a founder with Singh's track record bets 450 million on a different path, it's worth understanding why. I wrote a detailed breakdown of their approach, how it compares to what's becoming standard, and the questions worth asking. Read it here: https://t.co/K2VpSySn42 What do you think? Are you watching what's happening in robotics foundation models? Have you seen this latency problem show up in practice? #Robotics #Manufacturing #AI #FoundationModels #PhysicalAI #ArtificialIntelligence