Integrative Learning for Robotic Systems with Aaron Ames - #87

Reactive vs. Learned

Aaron Ames, a professor at Caltech, discusses the distinction between reactive and learned behaviors in robotics. He explains that reactive behaviors, like those seen in Boston Dynamics robots, allow robots to adapt to different terrains without learning new behaviors. Ames emphasizes that learning plays a crucial role in adapting to unforeseen environments, such as walking on sand or dirt, where models are complex and computationally intensive 1 2.

The deciding factor here is friction. So as long as it has sufficient friction when the foot touches down, you can do the same behavior you do on firm ground.

--- Aaron Ames

He highlights the use of Gaussian processes to update models based on new data, enabling robots to adapt to granular terrains 2.

   

Unifying Insights

Ames advocates for a unified approach in robotics, combining learning, control, and dynamics to create more advanced systems. He argues against relying solely on learning models, suggesting that integrating physics, computation, and mechanical design can enhance robotic capabilities 3.

Unify, unify, unify. So that's my argument on the forefront.

--- Aaron Ames

Ames describes a data-driven modification approach, where data is combined with learning models to optimize robotic interactions with the environment, as demonstrated in experiments with simple hopping robots 4.