Keep pulling the thread on Jitendra Malik.
Jitendra Malik and his collaborators developed a technique for robot locomotion called Rapid Motor Adaptation (RMA).
The goal of Rapid Motor Adaptation (RMA) is to enable a robot to adapt its motor actions to changing terrain on a timescale of a second or less.
The core principle of Rapid Motor Adaptation (RMA) is using the discrepancy between a commanded action and the actual sensed motion to infer terrain properties and adjust the robot's behavior.
The same neural network policy developed with Rapid Motor Adaptation allows a robot dog to walk across diverse terrains like rocky riverbeds, stairs covered in leaves, and loose mud piles.
Jitendra Malik's research group found it necessary to add vision to their robot to enable it to climb up stairs, a task the blind version of their system could not handle.
Jitendra Malik believes that quadrupedal (four-legged) locomotion is now a problem that is "pretty close to solved."
Jitendra Malik asserts that bipedal (two-legged) locomotion is significantly harder than quadrupedal locomotion and is still not a solved problem.
The breakthrough for deep learning in computer vision occurred when a model developed by Alex Krizhevsky and Ilya Sutskever, students of Geoff Hinton, significantly outperformed classical methods on the ImageNet challenge.
According to fossil data, the evolution of the human brain followed the evolution of the hand, particularly the development of an opposable thumb.
GPT-4 is capable of performing at the 90th percentile level on a law exam.
Jitendra Malik believes that AI will not be fully realized until it masters basic sensory-motor competencies such as seeing, moving, manipulating objects, and planning.
Jitendra Malik believes that language models will become more robust and less brittle if they are built on a foundation of physical interaction abilities.