Controls
Presets
Parameters
Quality
Camera
Model checks
The science
Idea. A control barrier function (CBF) turns "never touch an obstacle" into one linear inequality on the velocity command. For a point robot p' = u and a disc obstacle with centre c and effective radius R (obstacle plus robot), the barrier is h(p) = |p - c|2 - R2; h >= 0 is the safe set. The filter keeps dh/dt = a.u >= -alpha h with a = 2(p - c) (Ames et al. 2017, the standard CBF condition with the linear class-K function alpha h).
The QP. The safety filter edits the nominal command as little as possible: u* = argmin |u - unom|2 subject to the barrier inequality of every obstacle. With one active constraint the answer is closed form, u* = unom + max(0, -(a.unom + alpha h)/|a|2) a, a projection onto a half-plane. With several obstacles this page solves the 2D QP exactly by enumerating the projection onto each line and each pairwise intersection and keeping the cheapest feasible candidate. Because u = 0 always satisfies the condition while h >= 0, the QP is never infeasible inside the safe set.
Forward invariance. The condition gives h' >= -alpha h, hence h(t) >= h(0) e-alpha t >= 0: a robot that starts safe stays safe. The checks verify this on the Euler discretisation, where hk+1 = hk + a.u dt + |u|2 dt2 >= (1 - alpha dt) hk exactly. A larger alpha lets the robot get closer before the filter reacts (less conservative, sharper turns); a small alpha brakes early.
Simplifications and known limits. Single-integrator kinematics (velocity is commanded directly; real robots need a higher-relative-degree CBF, inertia and actuator limits), perfect position knowledge, circular static obstacles, no noise or delay. The filter is local: when the goal lies exactly behind an obstacle the robot slides onto the barrier where unom is anti-parallel to a and stops, a well-known CBF deadlock (the verdict readout reports it). The arena layout, gains and radii are illustrative, not taken from a particular robot.