Neural Lander

Scene

Manual: โ†‘/โ†“ throttle, โ†/โ†’ tilt, Space cut engine. Tip: use 5ร— and start braking right away.

Control loops

The same controller as a classic block diagram, running on the same sensors as the network. Wires glow with the signal they carry (blue positive, orange negative); red marks a limit that is active, dashed a gate that is off. Hover over a block to ring the neurons that implement it, or over a neuron to light up its blocks.

Mission profileterrain ยท glide profile ยท trajectory

Mission timeโ€”
Distance to padโ€”
Altitude / radarโ€”
Fuelโ€”
Horizontal speedโ€”
Vertical speedโ€”
Tilt (cmd)โ€”
Throttleโ€”
What is this?

The lander starts in lunar orbit at about 15 km, moving at about 1.6 km/s. It is flown by a small neural network whose weights were never trained. Each neuron was written by hand to compute one named quantity: "how far short of the pad we are", "safe approach speed for this distance", "too fast for this altitude", and so on. The weights follow from those definitions.

  • v1: the first design, frozen before it ever flew. It landed 84% of 3,000 random terrains.
  • v2: v1 plus reasoned fixes after reading v1's failed flights (each listed in designs/v2.js). It landed 3,000 of 3,000.

Hover over (or tap) a neuron on the right to see what it means, its live value, and its strongest inputs. Edges glow with the signal currently flowing through them: blue is positive, orange is negative.

The weights were compiled from a control design, so the same controller can be drawn as classic control loops. The Control loops panel runs that plain-code version on the same sensors and shows it next to the network.

Nudge (or press N) kicks the lander with a random velocity change, sized to the flight phase: up to 40 m/s per axis above 2 km of radar altitude, 10 m/s down to 100 m, and 3 m/s below that. Nothing is recorded: the network sees the new state on the next tick and has to correct. Red dots mark where each nudge happened.

Network