Physical AI Lab // Episode 2.7 // Quiz
Safety filter: control barrier functions
Bend a dangerous action back into the safe zone before it goes out
01
In the notebook you turn the filter off and the h curve drops below the zero line. What does that mean?
The robot is getting closer to the goal
It has crossed into the safety radius
The filter is bending the action back
The safe set has grown, so the constraint is looser
02
What exactly does the safety filter do at each step?
Find the safe action closest to the nominal one
Retrain the policy so it learns to avoid obstacles
Stop the robot at once whenever an action is unsafe
Scale the nominal action down by a fixed ratio
03
In open space, far from the obstacle, the robot's actions are almost identical with the filter on or off. Is that normal?
No, it means the filter isn't working
No, it means alpha is set too small
Yes, because h is negative far away, so the constraint is inactive
Yes: minimally invasive, it barely changes actions far away
04
The filter never failed in simulation, but on the real robot it still brushes the obstacle. What is the most likely cause?
The quadratic program is too slow: 0.4 ms can't keep up with 1 kHz
The distance or dynamics model is off, so the guarantee weakens
alpha is too large, making the filter too conservative
The filter can only handle one rule at a time
05
If the control input changes from velocity to acceleration, the original CBF condition can no longer be used directly. Why?
Acceleration control doesn't need a safety constraint at all
A quadratic program can't handle acceleration
Higher relative degree; it needs a higher-order method
Actuator saturation makes h negative
Check