The Humanoid Hub@TheHumanoidHub
Ashok Elluswamy, Tesla's AI lead, during a GTC discussion, highlighting the fundamental similarity in AI approaches for self-driving cars and humanoid robots:
- Hierarchical decision making is useful, but it has to be done as part of the same decision-making process as lower-level controls.
- We haven't seen the long tail of humanoid robotics, but Tesla has seen the long tail of self-driving, where high and low-level decisions have to be jointly made at a pretty high framerate.
- Optimus's architecture is designed in a similar way, where there's a hierarchy but it's all running as part of the same model and the latencies involved in decision making are well modeled.
- This architecture will scale quite well with humanoid robots.
- The distinction of the decision-making levels is only in the developer's mind. For the model, it's a continuous space of decision making, where there are dials available to make them more fine or coarse.
- Humanoids have more sensor modalities and higher degrees of freedom compared to self-driving, but the fundamental constraints remain the same: you need to make real-time decisions. There's obviously a hierarchy to these control signal outputs, but the lowest frequency cannot be too low, because the safety of the robot cannot depend upon things running at very low frequencies.
