Ming-Yu Liu
A research director at NVIDIA focusing on computer vision and generative AI, and a lead contributor to the Cosmos world model project.
Core Facts
Timeline (last 90 days)
Physical AI is far more fragmented than language, and there may not be a few unified foundation models; instead, several open models for specific domains will coexist.
The current bottleneck in physical AI is concentrated in data rather than compute: interaction data with physical annotations is extremely scarce.
As long as the policy success rate rankings measured by the neural simulator are consistent with those in the real world, it is possible to reduce the number of checkpoints requiring real-world testing and accelerate the pace of development.
Liu believes current world models are already good enough for navigation tasks, and that the real challenge lies in manipulation tasks involving contact, buffering, and deformation.
Liu defines world models as a collection of useful tools, drawing an analogy to the fact that the definition of AGI has yet to reach consensus.
Liu believes that safety is more critical for humanoid robots entering homes than for autonomous driving, because robots will be around children and pets.
Liu believes that neural simulators are more suitable in the early stages for policy validation rather than for directly training policies.
Tasks that are mechanically repetitive and time-consuming should be delegated to AI, while tasks requiring deep understanding, shaping judgment, and building core competencies should be handled personally.
All Facts (8)
Tasks that are mechanically repetitive and time-consuming should be delegated to AI, while tasks requiring deep understanding, shaping judgment, and building core competencies should be handled personally.
75%UnverifiedThe current bottleneck in physical AI is concentrated in data rather than compute: interaction data with physical annotations is extremely scarce.
50%UnverifiedAs long as the policy success rate rankings measured by the neural simulator are consistent with those in the real world, it is possible to reduce the number of checkpoints requiring real-world testing and accelerate the pace of development.
50%UnverifiedLiu believes current world models are already good enough for navigation tasks, and that the real challenge lies in manipulation tasks involving contact, buffering, and deformation.
50%UnverifiedLiu defines world models as a collection of useful tools, drawing an analogy to the fact that the definition of AGI has yet to reach consensus.
50%UnverifiedLiu believes that safety is more critical for humanoid robots entering homes than for autonomous driving, because robots will be around children and pets.
50%UnverifiedLiu believes that neural simulators are more suitable in the early stages for policy validation rather than for directly training policies.
50%UnverifiedPhysical AI is far more fragmented than language, and there may not be a few unified foundation models; instead, several open models for specific domains will coexist.
50%