We built a Rocket League coach that reads the whole field.
Most Rocket League stat sites count things: shots, saves, boost collected, demos. Counting tells you the score. It does not tell you why you lost. We built Replay Bench to answer the second question — and to do it well, the engine has to look at every car, on every frame, for the entire match.
A curated corpus. Serious horsepower. One engine that reads the whole sport.
The training corpus is hand-selected: 100,000+ pro and ranked Rocket League replays, stratified by tier, weighted toward the possessions that decide matches. Curation is the moat — the engine is only as good as the tape we point it at.
The engine reads the whole field on every frame — every car, every touch, every possession. That's expensive to do well, and we pay for it. The version of the engine you score against next month will be better than the one you score against today.
The output is a per-touch Threat score: 0 means harmless, 1 means a goal. When you take a 0.04-Threat shot and a teammate had a 0.31-Threat opportunity behind you, the engine sees both — and writes about it in the coaching note.
- 100K+Replays in corpus
- Per-frameThreat scoring
- Whole-fieldCoverage on every play
- All ranksFrom Bronze to RLCS pro
- 20sReplay → coaching note
- ProprietaryEngine architecture
A score that's only about your touch is a stat. A score that knows what your teammate had open is a coach.
In football analytics, expected goals tells you the probability a shot ends up in the net. It works because football is a slow sport with a small number of shots per match.
Rocket League is the opposite: thousands of touches per match, fast camera moves, six cars all moving at the same time. A touch-only score — one that grades a shot based purely on where the ball was — misses the entire point of the sport. The question isn't whether the shot was good. The question is: was it the right shot to take, given what else was happening?
Threat answers that. The engine evaluates every touch against the alternative possessions available at that exact frame. A 0.18 corner shot looks fine in isolation; it looks terrible if your teammate had a 0.42 backboard read open behind you. The engine sees both. The coaching note tells you about it.
Upload → parse → Threat → coach.
The full path from a .replay file in your collection to a coaching note in your browser, in roughly 20 seconds.
- Step 01
Replay parse
Native-format replay decoded frame-by-frame. The parse runs in your browser — your replay never leaves the client during this step.
- Step 02
Possession segmentation
Touches are stitched into possession episodes by clear-of-zone events. The engine treats each possession as a unit with a value.
- Step 03
Threat scoring
Every touch on every car gets a Threat score. The engine reads the whole field, every frame.
- Step 04
Coaching notes
Per-possession Threat deltas are converted into plain-English notes. Written from your seat in the match — what you saw, what you missed.
Four principles, no exceptions.
The engine has to be honest
We don't grade shots against an idealised pro. We grade them against the alternative possessions that were available at that frame. If you took the best shot on the field, the note says so.
Coaching notes are written, not auto-generated
The engine produces structured per-possession output. We have a tight library of human-written templates that fit the output to a sentence — no LLM hallucinations, no generic 'try harder' notes.
We rank-stratify the analysis
Coaching a Plat player like a GC will alienate them. Every coaching note is filtered against the rank tier it's written for — see our per-rank coaching guides for what this means in practice.
Engine ships keep getting better
We're training on a moving corpus. The version of the engine you score against next month is built from a better corpus, on better hardware, against better evaluation tape. You don't have to do anything to opt in.
See what the engine sees.
Join the Discord for early access and to talk to the team.