Good questions.

Kiari chooses how much to apply each of five controls: throttle, steer, pitch, yaw and roll. Each one is an analog axis from −1 to +1 rather than an on/off button.

Jump, boost and handbrake remain on/off buttons. Calling Kiari a continuous-action bot describes its analog axes; saying every car input is analog would be inaccurate. The interactive control diagram on the landing page shows both kinds.

Parameters are learned numbers that shape a neural network’s calculations. Kiari’s complete training model has 11,924,502: 3,488,768 in a shared trunk, 4,228,117 in the policy and 4,207,617 in the critic.

The trunk reads the 323 observation values through four layers of 1,024 units. The policy and the critic each add four more layers of 1,024 units on top of it. The policy turns what the trunk sees into driving controls; the critic predicts future training return and helps improve the policy. Only the policy drives the car, and both of them train the trunk.

The 323 values describe the ball, its own car, the opponent, the boost pads and its last input, all read from the blue side of the field. None of them predicts where the ball will go; that has to be learned.

The policy turns what the shared trunk has read into the eight controls at once: a value between −1 and +1 for each of the five analog axes, and on or off for each of the three buttons. Sampling during training lets it explore different choices.

In deterministic inference it takes its most likely choice for every control.

It makes 20 decisions per simulated second: one every six physics ticks at 120 Hz. With the five-tick action delay, the old control continues while the new decision is pending, and the new one starts on the final tick of that cycle.

The latest goal share and its 95% confidence interval are shown with the benchmark on the landing page. A share above 50% means Kiari scored more goals than Nexto in that series.

That result belongs to the named checkpoint and protocol. A different opponent, match format or set of starting situations asks a different question.

The landing page lists the mechanics the trainer has counted, with their current rates: air dribbles and goals from them, wall dashes, wavedashes, zap dashes, ground carries, flip resets and shots with a reset flip. None of them is scripted; each one appears in self-play and the training rewards keep it.

Kiari V4 starts with the basics: goals, moving to the ball, facing it, air time, strong touches, the ball toward goal, boost, bumps, demos and the kickoff. Air dribbles, wavedashes, flip resets and the HyperNoStack mechanic rewards are off at the start. A dedicated air-dribble reward, flying the ball to goal and keeping it high, is planned once Kiari reaches Necto level. Most mechanics are counted by the reward that trains them, so the list starts short and grows.

No. Goal share is Kiari’s goals divided by both sides’ total goals. The headline is the newest completed series of Kiari V4’s repeated test against Nexto, with its goal target named beside it; it does not report a match win rate.

An arena match has its own score. That score and the training benchmark are separate measurements.

A finite evaluation has uncertainty. The 95% confidence interval shows the statistical uncertainty around the measured goal share; a small change inside that range is weak evidence of improvement.

The headline protocol is random starts (random car, ball and velocity states, no kickoffs), deterministic decisions, 30,000 goals, 4,096 arenas played from both sides, seed 20260725. The goal share depends on the start states and on the arena count, so each protocol is its own series: a test measured any other way stays in its own tab of the test history and never joins the climb. Necto is tested the same way, also to 30,000 goals.

The interval does not account for every source of variation or transfer the result to human matchmaking. Repeated comparable evaluations give a better view than a single score.

The live counter and training readings come from the project’s telemetry and refresh every twenty seconds. Between two readings the counter moves at the measured rate toward the newest sample, never past it. The headline goal share comes from a completed evaluation and names the checkpoint it tested.

The model’s 11,924,502 parameters are calculated from its implemented architecture and checked against the training configuration whenever the site’s tests run. Those documented facts are distinct from changing live readings.

Artwork, the 3D signal trace and the control demonstration are illustrations. Missing telemetry is labelled “not reporting” rather than replaced with a made-up number. The mechanic rates are counted by the trainer itself, per player and minute of simulated play.

One step is one car decision collected in a simulated environment. The total accumulates across training arenas; it does not count matches, gradient updates or real-world seconds.

The headline rounds the latest training count to the nearest 50 billion steps (≈), or to the nearest billion while it is below 50 billion. The exact sampled count is shown underneath. Both refresh every twenty seconds, alongside current steps per second for the whole training loop. An evaluated checkpoint can be older than the current training state.

Kiari learns through self-play: it plays against its current self, and PPO uses those games and the training rewards to update the policy. It never trains against older versions of itself; games against older checkpoints are only used to measure it.

Human replays are not its training data. Nexto and Necto are evaluation opponents, not training opponents.

Necto is a community Rocket League bot that came before Nexto. Kiari plays it with the same protocol as Nexto: random starts, deterministic decisions on 4,096 arenas from both sides, to 30,000 goals per test.

It is the second yardstick: it shows progress while the gap to Nexto is still large. Its results are listed separately and never mixed into Nexto’s numbers.

Supersonic Legend (SSL) is the project’s target. Kiari V4 has not been placed on the human matchmaking ladder, and no rank is claimed for it.

The benchmarks measure goal share against Nexto and Necto under random starts / deterministic / 4,096 arenas / 30,000 goals / seed 20260725, with a 95% confidence interval. They do not convert into MMR. A rank needs games against ranked human opponents.

Try the browser arena. The outcome depends on the matchup, and no benchmark can predict an individual match.

The public benchmark gives a repeatable comparison against Nexto. Your arena match is a separate, playable encounter.

Yes. Open arena.kiaribot.com to watch or play in your browser. The game is streamed, so you do not install the bot.

One visitor drives at a time. Other connected visitors watch and wait in the queue. The arena reports its current capacity and availability.

When your driving turn ends, the next eligible viewer is promoted. You can remain connected to watch and wait for another turn.

Closing the connection gives up your current place. The arena shows the active role and queue state.

The browser arena is the public way to try Kiari. For local availability, contact [email protected]. This site does not promise a download, price, release date or update entitlement.

Running a saved policy and training a new one are different things. The site does not distribute the training recipe or private model weights.

Kiari V4 is Kiari’s fourth version and the only one this site follows. It started from scratch on 16 October 2026 on GigaLearnFast, a training engine that runs PPO on fused CUDA kernels, with a shared trunk and two heads of four layers of 1,024 units reading a 323-number observation. It uses the same network, observation and controls as V3, and it begins with rewards for the basics only.

Earlier generations, Kiaribot V2 included, and the earlier V3 runs are retired, and none of their tests is mixed into this record. Every number on the landing page belongs to this run: the live readout, the scores against Nexto and Necto, every test in the climb and the mechanics it has learned.

More training creates opportunities to improve; it does not guarantee progress. Compare the completed evaluations on the learning curve, keeping the protocol and confidence intervals in view.

The training counter measures experience collected, not a promise of strength. A future checkpoint still needs its own evaluation.

Kiari is an independently developed project. Its public site reports the training and evaluations that are available, rather than implying a larger controlled study.

A change in one run is not automatically evidence that the change caused an improvement. Comparable evaluations and their uncertainty matter.

The policy keeps the parameters it has already learned. Pausing training does not erase those abilities, and a saved checkpoint can still be used for inference.

If telemetry reports a stopped process, the site can show that state and its last readings. If telemetry cannot be reached, it shows “not reporting”. Arena availability is separate.

The reward recipe, learned weights and private training configuration are not published.

The site shares the model’s dimensions, control format, selected training measurements, curriculum emphasis and available evaluations. Its data reader publishes only those selected fields.

These are project-reported measurements, not an independent certification. The site identifies the tested checkpoint, protocol and uncertainty so you can judge what each result supports.

The arena lets you encounter an available policy snapshot in a live game. It is not a recording, but its snapshot can differ from the latest training or evaluated checkpoint.

KiariBot is an independent project and is not affiliated with Psyonix or Epic Games. Training takes place in a physics simulator. Rocket League, its car designs and rank emblems belong to their respective owners.