See the algorithms think.
Six hands-on simulations covering estimation, kinematics and control. Drag, slide and step through each one, and watch the maths update with you.
SLAM
EKF-SLAM in 2-D. Drive a robot, watch uncertainty grow and then collapse at loop closure.
Unit IVKalman filter
Blend a prediction with a measurement, then track a moving target in real time.
Unit IVMonte Carlo localization
The exam's four-particle corridor problem, step by step, plus a free-running particle filter.
Unit IDH transforms
Build a link transform one operation at a time in 3-D, and drive a two-link arm.
Unit IIRouth–Hurwitz
Type any characteristic polynomial and get the full Routh array with working and roots.
Unit IISecond-order response
Drag the poles and see overshoot, settling time and a robot joint respond.
EKF-SLAM: map the world while finding yourself in it
The navy robot is the truth. The teal robot and landmarks are the EKF-SLAM estimate with 2σ ellipses. The orange ghost uses odometry alone. Watch the ellipses grow as the robot explores, then shrink when it sees the start again (loop closure).
1 · Predict
Each step the robot moves with noisy wheels. The EKF pushes the pose through the motion model and adds process noise, so the robot ellipse grows.
2 · Observe
Landmarks within range return a noisy range and bearing. A new landmark is added to the state vector with its uncertainty, which is tied to the robot's.
3 · Update
Known landmarks correct the robot and every correlated landmark. When an old landmark is seen again, the whole map snaps tighter: loop closure.
Kalman filter: one step, then tracking
A Kalman filter keeps a Gaussian belief. Predict moves it and makes it wider; update blends it with a measurement and makes it narrower. The defaults below are the worked example from the slides.
Constant-velocity Kalman filter
Orange dots are noisy position readings. The teal line is the filter's estimate, with a shaded 2σ band. Raise Q to make the filter trust motion less and react faster; raise R to make it smooth harder.
Monte Carlo localization in a corridor
Each particle is a guess of where the robot is. Particles that explain the sensor reading well get more weight, then resampling copies the good ones and drops the bad ones.
Denavit–Hartenberg: build a link in four moves
ⁱ⁻¹Tᵢ = Rot(z, θ) · Trans(z, d) · Trans(x, a) · Rot(x, α). Scrub the build slider to apply the four operations one at a time, and drag to orbit the 3-D view.
| cθ | −sθ·cα | sθ·sα | a·cθ |
| sθ | cθ·cα | −cθ·sα | a·sθ |
| 0 | sα | cα | d |
| 0 | 0 | 0 | 1 |
Top-left 3×3 block: rotation, whose columns are the new x, y, z axes written in the old frame. Last column: position of the new origin.
Two-link planar arm
Move the sliders for forward kinematics, or drag the orange gripper for inverse kinematics. The shaded ring is the reachable workspace.
Routh–Hurwitz stability checker
Enter a characteristic polynomial. The Routh array, every calculation, and the root locations update live. Each sign change in the first column is one root in the right half-plane.
Second-order response of a robot joint
T(s) = ωn² / (s² + 2ζωn·s + ωn²). Drag the poles in the s-plane or use the sliders. The default is Q6: T(s) = 25 / (s² + 6s + 25).