Kiari V4Rocket League · self-play from scratch
A Rocket League bot that learns by playing against itself. Version 3.1 started from zero on 16 October 2026: no scripted plays, no copied games. Every touch is a decision it learned.
Training steps
5B
5,002,346,496
car decisions collected since 16 October · sampled every 20 s
Steps per second
not reporting
whole training loop
Goal share vs Nexto
not reporting
latest completed test
Trainer
Paused
run Kiari V4
Original fan artwork · an illustration, not gameplay
01 — The benchmark
Nexto is one of the strongest community Rocket League bots. Kiari plays it in many arenas at once, on both sides of the field, until the goal target is reached, then the goals are counted. This is the latest completed series.
No test against Nexto has finished yet for this run.
And against Necto
Necto is the community bot that came before Nexto. The same kind of test, measured on its own. It is the second yardstick, never mixed into Nexto's numbers.
No test against Necto has finished yet for this run.
02 — The climb
Every point is one completed Kiari V4 series against Nexto, placed at the training step it measured. A line keeps one protocol: the same start states, arenas, seed and goal target every time. When the test changes, a new line starts and the old one stays behind it, dimmed.
The evaluation record could not be read just now.
Newest first. Search a checkpoint like 2.50B or a date like 2026-10-16.
Other protocols ask different questions (another opponent, goal target or setup). They are kept here and never mixed into the climb.
The evaluation record could not be read just now.
03 — What it learned
No mechanic is scripted. Each one is a habit the bot finds in self-play and its training rewards keep. Kiari V4 starts with the basics; air dribbles, wavedashes and flip resets are off at the start, and a dedicated air-dribble reward is planned once Kiari reaches Necto level. A mechanic appears here once the trainer has counted it: the line runs from its first step to now, the number is its current rate, per player, in simulated play.
Nothing counted yet. The mechanics appear here as the trainer counts them.
04 — The play
01 / 03
The ball is rising. The opponent is closing. Twenty times a second, Kiari reads 323 numbers about the ball, both cars, the boost pads and its own last input.
Every input is computed from the game state, the same way in training and in a match. The artwork is an illustration, not its view.
02 / 03
A small correction changes the shot. The numbers pass through a shared trunk of four layers of 1,024 units, then four more layers of the policy, which settles on how much to steer, accelerate and rotate.
There is no scripted shot sequence. Every control is chosen by the learned network.
03 / 03
Throttle, steer, pitch, yaw and roll are analog controls, set between −1 and +1. Jump, boost and handbrake are on or off. Then it reads the field again.
Twenty decisions per simulated second, on 120 Hz physics.
05 — In training
Kiari V4 trains on GigaLearnFast, a training engine that runs PPO on fused CUDA kernels. It started from scratch on 16 October 2026: a shared trunk and two heads of four layers of 1,024 units, reading 323 numbers from the match. Every episode starts from a situation; each band is the share of new episodes that start from it, read live from the run's configuration.
60%Open play
Shares are emphasis on starting situations, not time on a clock.
The whole loop: collection and learning together.
Steps gathered per second by the simulated arenas.
Batches of experience collected and learned from.
Exploration across throttle, steer, pitch, yaw and roll.
Exploration across jump, boost and handbrake.
How far each update moves the policy. Kept small on purpose.
How well the critic predicts the return it is about to see.
Since the trainer process last started.
06 — Inside Kiari
A shared trunk reads the 323 numbers. On top of it the policy drives and the critic judges the position; the critic helps it learn and never touches the controls.
Drag to turn · arrow keys work too
One point per unit of the published topology. The signal is an illustration; this view contains no learned weights or live activations.
323 numbers, twenty times a second. Every car reads the field from the blue side, so a car on orange sees the same game mirrored. There is no ball prediction among them: where the ball is going is something Kiari has to learn.
The owner's own observation layout, live since 14 October 2026
The shared trunk
3,488,768parameters
4 x 1,024
Four layers of 1,024 units read all 323 numbers first. The policy and the critic both build on what it learns, and both of them train it.
Linear, LayerNorm, ReLU, four times. Gradients from the policy and from the critic both reach it.
The policy
4,228,117parameters
4 x 1,024
Four more layers on top of the trunk, then the controls. This is the part that drives.
Linear, LayerNorm, ReLU, four times, then the eight controls. With the trunk, 7,716,885 parameters lie between the inputs and the controls.
The critic
4,207,617parameters
4 x 1,024
Its own four layers on the same trunk. It answers a different question, how good is this position, and nothing it computes reaches the car.
It ends in one value. During training it also teaches the trunk what matters for the result.
What comes out
The policy ends in five continuous control values and three binary buttons.
Every decision sets each analog axis between −1 and +1 and each button on or off. Deterministic play takes the policy's most likely choice for every control.
Try continuous control
Slide an axis. A value of +0.37 is a different instruction from +1.00. The three buttons have just two states.
An interactive explanation, not a live policy output. These controls only change this drawing.
The policy reads what the trunk has computed and sets all eight controls at once: a value between −1 and +1 for each of the five analog axes, on or off for each of the three buttons.
During training it samples its choices, which lets it explore. Deterministic play takes its most likely choice for every control.
The policy decides every 6 physics ticks at 120 Hz: 20 decisions per simulated second. A 5-tick action delay keeps the previous control in place while the next decision is pending.
07 — The level
SSL is the target. Kiari V4 is measured against bots, not people: its place on the human ladder has not been tested yet.
Bronze
Tier 1 of 8 on the competitive ladder. The badge shown is Bronze III.
Silver
Tier 2 of 8 on the competitive ladder. The badge shown is Silver III.
Gold
Tier 3 of 8 on the competitive ladder. The badge shown is Gold III.
Platinum
Tier 4 of 8 on the competitive ladder. The badge shown is Platinum III.
Diamond
Tier 5 of 8 on the competitive ladder. The badge shown is Diamond III.
Champion
Tier 6 of 8 on the competitive ladder. The badge shown is Champion III.
Grand Champion
Tier 7 of 8 on the competitive ladder. The badge shown is Grand Champion III.
Supersonic Legend
The highest competitive tier and Kiari's target. It has not been placed on the human ladder.
Target · not a measured or verified competitive rank
What the evidence says. The Nexto and Necto series measure goals in a fixed bot protocol. They do not measure human matchmaking or convert into MMR. A rank claim needs games against ranked human opponents.
Rank emblems © Psyonix / Epic Games. Not affiliated with or endorsed by Epic Games. Rocket League rank guide
08 — The arena
Watch Kiari in the browser arena, or join the driving queue and put your own instincts to the test.
Free browser arena. Availability and controls are shown there.