Keep pulling the thread on Jim Fan.
An AI won a gold medal in the International Mathematical Olympiad (IMO) in the current year.
AI technology was instrumental in work that led to a Nobel Prize for protein folding.
The "physical Turing test," where a robot's domestic work is indistinguishable from a human's, is the next or perhaps last grand challenge for AI.
Video generative models like Sora and Veo are data-driven systems that compress both rendering and a physics engine into a single neural network.
The Groove2Dreams model was created by fine-tuning a video world model, pre-trained on hundreds of millions of hours of YouTube videos, with a small amount of in-house teleoperation data.
NVIDIA's guiding philosophy for robotics development is a "data maximalist" and "model minimalist" principle.
The GROOT-N1 and N1.5 models use a Vision Language Model (VLM) backbone and a diffusion transformer to output a float vector for robot actions.
All future versions of NVIDIA's GROOT models will be open source.
The probability of developing a robot significantly more capable than Rambo, the golf-cart-driving ape, by the year 2040 is close to 100%.
Data for training humanoid robots, such as motor control actions, cannot be scraped from sources like Wikipedia and must be collected directly.
The Stanford Aloha project uses static bimanual grippers that are more robust and easier to replace than humanoid robots, enabling continuous data collection for long-horizon tasks.
The Dex-Umi method uses a mechanical exoskeleton that mirrors a robot's joints and constraints to collect data directly, and can be applied to different dexterous hands like the Inspire hand and the X hand.