AI solves: every map raced to its finish, then made faster
What found, settled and capped mean
A solve is one whole race of WRAI 1.1 (the same network for the whole queue) from the map’s real start to its finish, in the exact simulator, in one endless turn: fall damage 50, the parachute on, never put back. Every move is the network’s own: Fever races many copies of it every 5 seconds and keeps the best few going; where the worm is stuck, a lucid dream tunes it at that spot for a while, then Fever again. A solve is never a world record: a real world record always wins over it.
Found: a first race to the finish. Settled: from the exact moment each 5-second stretch of the race starts, on two grids 2.5 s apart so that no 5-second stretch falls between them, fresh copies raced that stretch at rising temperatures (at least 504 of them on the lean profile, 2,296 on standard, 5,880 on thorough), and none of the ones that got ahead led to an earlier finish once the rest of the race was raced again. A faster race only counts when it finishes at least one decision (2 frames) sooner, and it is replayed exactly before it is saved. So settled means nothing faster was found with that many tries, not that it is the fastest possible. Capped: its time ran out before every stretch was settled; it carries on after the other maps. Not solved: the search was parked or failed; see its reason below. An exact partial is not a finish or proof of an honest route. BigRR maps are solved with the simulator’s rope, not WA’s Batty Rope.
The queue
Not solved
What it did last
A race, replayed
The most recent run
Progress
Never completed: the maps it has never beaten
How these are counted
Never completed: every map training races that it has not once finished from its real start since , in a training race or a test race. A map it finishes leaves the list at the next refresh: the list is worked out again at most once a minute.
How far it gets: no record of each map's best race from its real start is kept, so it shows one of two signs. Best seen (with a bar): the furthest of its real-start tries this site has seen since . One racer in each worker saves its latest try at a whole map, so this is a sample: the real best may be further. When no try has been seen yet, its first stuck place (words only, no bar): the stuck place nearest its start (the Stuck spots rules: failing there at least 20 times, in at least a quarter of the races that got there), counted since from every race, practice starts too. This is only a rough sign: practice starts near the finish can put it close to the finish while races from the real start get nowhere near it, so on a long course it can look much closer than it is. Closest to finishing lists the maps with a try seen first, then those with only a first stuck place, each group by the number it shows. Older heat map: its heat map was rebuilt after these places were counted, so they can be off.
Tries: a RopeGP map's training races from its real start since (the curriculum's own count, exact). A kept generated map's running count, when the trainer keeps one (halved at 200 races). Other maps have no counter: about is the most races that crossed any 100 px of its route since , practice starts included, so it is often 1 to 2 times the real number, and more on a long course. Most tries sorts these estimates together with the exact counts.
Stuck spots: its stuck places and its issues (stuck places with memories); Open shows them on its map in the Maps tab. Heat map: v2 is made by the current rules, v1 by older ones; v2 refused: under the current rules there is no way from its start to its finish (a gap too tight for the worm), so its v1 heat map is kept. Approval: whether its start, finish and walls are approved. There are no pictures in the list: open a map to see it.
Held-out maps never finished in a test
Where it keeps getting stuck
Memory: the issues it has solved
How WRAI 1.1 sees and remembers
Ray sight. 64 rays all the way round the worm and one along its aim (where a rope shot would go), each stopping at the first terrain pixel, exact to the pixel's edge, or at the water. The pictures show the map in 4–64 px squares; the rays give the borders exactly, with how good the spot each ray reaches is on the heat map.
Flight arc. Where the worm would fly from its speed right now with nothing holding it: W:A's own free flight (gravity 0.24 px per tick², no drag), followed for 3 s until it would land on terrain or in the water. On the rope it shows where letting go now would send it, so a pump's extra speed stretches the arc; it only flies that path once nothing holds it.
Issues. The places it kept getting stuck (the Stuck spots tab's tally), each with a number that never changes. Training replays them: those races start 150–700 px of route before the place, and a race whose worm gets into the issue’s finish (a circle past the place) without drowning passes it. The finish is set automatically at the end of the world record’s run through the place, or by you in its Details; an issue with no finish is on hold (not replayed, no memories) until one is marked on the Needs a finish tab. With a world record’s run, an issue gets its own course: a finish where the record’s run ends (or yours), its replays starting where the record’s run started (its speed and rope too), and a heat map just for that section leading to the finish -- the race ends there.
Memories. Only from those replays (the issue training windows, where the issue is known exactly). Each issue keeps up to 8: every pass while there is room, until the issue is mastered (90% of its last 50 tries passed); after that only a pass faster than its slowest memory gets in, replacing it. A memory’s time runs from its course’s start to the moment its worm first gets into the finish circle from the course’s side (through a wall does not count), so their times compare fairly. Your runs are timed from their moment nearest the start. A memory kept from before its start moved is marked other start: its time is not compared, and the next pass replaces it first. Your memories (from your recorded runs) come first and are never replaced. One memory bank is shared by every racer.
Issue matcher. A helper network (1.2M parameters) trained only on those replay moments: which issue a moment looks like and how far before the place it is. When it is at least 30% sure, the racer is handed that issue's memories at that point: how soon each run got past, where it went over the next 3 s, how fast, the rope, and its next keys. It can recognise an issue on a map it has never seen, if the place looks alike.