It is still climbing.

279,118,765,702 training steps 146,155 steps per second

read off the trainer every twenty seconds

What the machine reports

3,510 PPO iterations Batches collected and learned from.
146,155 Steps per second Collection and learning together.
2.007 Continuous entropy Steering, throttle, air roll. Holding near 2.0.
0.850 Discrete entropy Jump, boost, handbrake. Roughly doubled to 0.85.
0.0038 KL divergence How far one update moves the policy.

A step is one decision by one car in one simulated match, and this machine makes about a hundred thousand of them a second. Both entropy readings are still wide open this deep into a run, which is the policy keeping its options rather than settling on one answer.

the axis is training steps, not dates

How it got here

Every percentage is a real series played against Nexto, at the step number the run was on when it happened.

  1. step 0

    A random policy

    Training starts from noise. Nothing in the code tells the car that the ball matters.

  2. 139.5B

    53.9% of goals

    The first series on the record, confidence interval 52.9 to 54.9. Ahead of Nexto, barely.

  3. 200.1B

    63.3% of goals

    Confidence interval 62.4 to 64.3. Nine and a half points gained over sixty billion steps.

  4. 205.0B

    A dead fixture, found and rebuilt

    An audit caught one training situation logging 643 bump attempts and no successes. Rebuilt, and back in the mix at three times its old weight.

  5. 260.0B

    69.8% of goals

    Confidence interval 68.9 to 70.7. Six and a half more points, and the curve has not flattened out.

Fitted across every series on the record, the goal share gains about a point per ten billion steps and the last thirty billion came in a little above that. The current reading is live further down this page.

from the trainer's own training config

What it is practising right now

The training mix is set by hand and changed when the evaluation says something has stopped paying.

  1. Kickoff

    24%

    Both cars on the line, ball dead centre. Every match starts here, so it takes the largest share.

    KickoffState Match play

  2. Scrambled play

    20%

    Cars and ball dropped anywhere at random speeds, so it has to read a situation instead of replaying one.

    RandomState Match play

  3. Air dribble

    17%

    Ball in the air ahead of the car: keep it balanced on the nose and fly it at the net.

    AirDribbleSetup Air control

  4. Coming off the wall

    13%

    Just off the side wall with the ball ahead and above. Steady first, swing later.

    WallDetachControlSetup Wall play

  5. Leave the shot, take the goalie

    7%

    The ball is already going in. Ignore it and remove the defender who is about to save it.

    AirBumpReleaseSetup Pressure and contact

  6. Ground into an air carry

    6%

    The piece people skip: lifting a bouncing ball off the floor and turning it into a carry.

    AirCarryBridgeSetup Air control

  7. Caught out with no boost

    4%

    Wrong side of the ball, empty tank, an attack coming. Rotate back rather than dive.

    LowBoostTransitionSetup Defending

  8. Carry, release, take the keeper

    3%

    Hold the air carry, release at goal, then turn into the defender and knock it out of the save.

    AirCarryBumpSetup Pressure and contact

  9. Shot on goal

    3%

    A shot arrives at its own net, slow roller to full power. Keep it out.

    SaveShotState Defending

  10. Air dribble into a flip reset

    2%

    Mid air dribble, flip upside down and get under the ball. It used to never attempt this.

    AirDribbleToResetSetup Air control

  11. Air dribble off the wall

    2%

    Car and ball both on the side wall. Peel off the surface and take the ball into open air.

    WallAirDribbleSetup Wall play

Shares are relative emphasis, normalised, not minutes on a clock. Below them are the trainer's own series, grouped, every axis in training steps. The deck counts itself off the board it read.

the measurement

Against Nexto

Nexto is a well known public bot, which is why it is the yardstick. Beating something other people can go and check is worth more than beating something only this machine has seen.

KiariBot

69.78%

95% confidence 68.88% to 70.68%

6,979 goals scored

Nexto

30.22%

3,022 goals scored

The two policies play until ten thousand goals have been scored between them, and the number is the fraction that were KiariBot's. Nothing on this machine keeps score by match, so no such figure exists here.

The band beside it is a 95% confidence interval, about nine tenths of a point either side. Some of the changes made to this run are worth less than that, which is why the band is printed every time the number is.

You

How you do against it has never been measured, and inventing a figure would be worse than saying so. The arena answers it in about a minute.

Take the wheel

the measurement

live
Goals in the series
10,001
Measured checkpoint
278,850,985,606
Run
gen5

Goals counted, never matches.

an estimate, with its reasoning shown

Roughly where that puts it

One opponent scores this run, so a place on the human ladder has to be inferred from a single number. Three links get there and the last one is an assumption, which is why this comes out as a range and not a pin.

69.78% of goals against Nexto 95% confidence 68.88% to 70.68%

the one number this rests on

  1. Where Nexto sits

    The community puts Nexto around Grand Champion. That is a consensus rather than a measurement, and every step after it inherits that.

  2. What the series says

    KiariBot takes most of the goals off Nexto over ten-thousand-goal series, and the figure above carries its confidence interval. This link is measured, and it is not close.

  3. The jump to a human ladder

    Goal share against one bot does not convert into a place among people, and nobody has measured that conversion. So the series puts a floor under this and no ceiling on it, and the range is drawn that way.

KiariBot Grand Champion to Supersonic Legend

  • KiariBot at least the tier Nexto is placed in, with no ceiling on it that this series can see
  • Nexto where the community generally places it. A consensus rather than a measurement, and the estimate starts from it.
  • The width is the honest part. One bot against one bot does not resolve a human rank any finer than this.

Defending is the thinnest part of its practice, about one part in sixteen, and people press differently from a bot. Patient play that makes it commit first is the seam. All of that is faster to check in the arena than to argue about.

arena.kiaribot.com

Take the wheel

The match is rendered on the machine that is training and streamed out of it. Eight people fit at once. First in line drives, the other seven watch.

the machine rendering this

146,155

training steps per second, while you drive

  1. 1 drives
  2. 2 next
  3. 3
  4. 4
  5. 5
  6. 6
  7. 7
  8. 8
Eight seats on one stream. The seat holding the wheel is whoever reached the front of the queue, and the other seven watch the same render. This row is the rule, not a count of who is in it.

When a turn ends the next viewer is promoted and told so, and whoever was driving goes to the back of the queue without losing the stream. One render process means one camera, so camera and match settings belong to whoever holds the wheel.

If nobody is driving, the wheel is free the moment you open it.

Find out if it can beat you

The arena is free and stays free, eight people at a time. If you want it without a queue, the paid tiers put the bot in your own copy of Rocket League: freeplay, private matches, custom playlists, a sparring partner that does not get tired of you.

Pricing is not published yet.