CBF Safety Filter

Key Insight

Wrapping a learned policy with a Control Barrier Function (CBF) safety filter guarantees physical safety without requiring the policy itself to learn complex constraint boundaries. The safety filter monitors the robot's state and projects the policy's raw control inputs onto a mathematically safe set of actions. If the learned policy proposes a command that would lead to a collision or joint limit violation, the filter minimally modifies the command to ensure safety while preserving the original policy's intent as much as possible.