Atlas trains with great behavioral models

in Popular STEM15 days ago

Atlas trains with great behavioral models



Image source:


Switching to the new Atlas improvements.


Boston Dynamics did not take the Atlas to the first robotics Olympics in China, for one reason, the team was connecting LBMs (Large Behavioral Models) to the robots. Instead of handwritten commands, Atlas learns patterns of human action from large data sets.




The goal is simple to say and difficult to achieve.


Make a humanoid truly versatile, able to adjust what it does based on what it finds, what it does so far, transfer objects between baskets, open, close, place items in moments, walk, bend, classify, organize and continue working even when someone tries to interrupt it.


It is not Olympic speed, it is more careful and stable execution with skill and force control in focus, the difference is in generalization, new skills that previously required weeks of programming, now come through demonstrations and data. without writing another wall of code.



Image source:


And how does it work?


Translated, the LBM is a large model trained with many human demonstrations, it learns long-range manipulation sequences, open, pull, rotate, position and then adapts them to variations of object, posture and scenario, instead of “if the object is like this, do this”, Atlas infers the gesture that the situation requires.


Result, less stubborn robot and more robot that understands the task, even when someone pushes the basket or changes the object's location, this in practice is, if humans promise to act in environments that already exist, home, warehouse, factory, the bottleneck is teaching them everything they will have to do, with LBMs, the pace of learning is accelerated, less demonstrations for more behaviors, integration with visual perception and control of the whole body and a shorter path between show me how it's done and let it be done, I'll do it.



Image source:


Incorporating a "behavioral brain"


This is where projects with research partners come in to provide data scale and rigor, yes, it is clear that it is still in the development phase, Atlas is still somewhat slow and deliberate, it needs to prove robustness on a daily basis, functional safety, in human coexistence, standardized performance evaluation and coverage with soft, reflective or confusing objects.


Questions also arise about computational cost, training and continuous updating and how to measure understanding beyond the flashy video, but the direction is clear, go off script and learn about the world like all of us.



References 1


Follow my publications with the latest in artificial intelligence, robotics and technology.
If you like to read about science, health and how to improve your life with science, I invite you to go to the previous publications.
You want to win, play HARRY-RAID