Zombie Resurrection

Rust workspace for turning raw RoboCup SSL logs into model-ready training samples.

What it provides

• Live mode: parse .log / .log.gz files with loguna, clean them, segment sequences, and emit per-robot training samples on demand.
• Archive mode: preprocess logs into compressed rkyv + zstd shards with manifest, split bundle, and normalization stats.
• Review tool: embedded Rust review server plus React/Vite web app for sequence review, playback, overlays, and bulk triage.
• Optional tch feature in zr-core: convert sample batches into weighted tch::Tensor values.

Workspace crates

• crates/zr-core - core library only
• crates/zr-cli - CLI tools
• crates/zr-review - embedded local review server
• apps/zr-review-web - review web frontend

Binaries

• zombie-resurrection — interactive preprocessing with optional manual review
• zr-auto-pipeline — fully-automated raw → training-data pipeline (no human intervention)
• zr-review — embedded local review server

Quick start

Generate a starter config:

cargo run -p zr-cli --bin zombie-resurrection -- init-config --output zr-config.json

Audit one log:

cargo run -p zr-cli --bin zombie-resurrection -- audit \
  --input /path/to/raw-log \
  --team-name YourTeam

Preprocess logs into archived shards:

cargo run -p zr-cli --bin zombie-resurrection -- preprocess \
  --input /path/to/raw-logs \
  --team-name YourTeam \
  --output-dir data/processed \
  --review-file data/review.json

Automated pipeline (no review, enhanced cleaning):

cargo run -p zr-cli --bin zr-auto-pipeline -- \
  --input /path/to/raw-logs \
  --team-name YourTeam \
  --output-dir data/auto \
  --verbose

Run the embedded review server:

cargo run -p zr-review --bin zr-review -- \
  --logs-dir /path/to/raw-logs \
  --team-name YourTeam \
  --review-file data/review.json

The frontend is built and embedded automatically by build.rs.

For frontend-only iteration you can still run:

cd apps/zr-review-web
npm install
npm run dev

Library usage

use std::path::PathBuf;

use zr_core::{LiveDataset, PipelineConfig, ReviewStore, TeamSelector};

let mut config = PipelineConfig::default();
config.target_team = TeamSelector::Name("YourTeam".to_string());

let dataset = LiveDataset::new(
    vec![PathBuf::from("/path/to/raw-log")],
    config,
    Some(ReviewStore::default()),
);

for sample in dataset.iter().take(8) {
    let sample = sample?;
    println!("{} -> {:?}", sample.metadata.game_id, sample.target);
}
# Ok::<(), anyhow::Error>(())

Current v1 scope

• Raw source: VisionTracker2020
• Training target: per-robot next relative delta (dx, dy, dtheta)
• Manual review: web queue, playback, scrubber, overlays, compare mode, keep/drop/needs-attention with notes
• Archive format: rkyv payload compressed with zstd
• Split export: train/validation/test assignments plus normalization stats and sample weights
• Identity handling: Hungarian-style matching with gating and suspicious-swap flags
• Live-play handling: referee commands plus motion heuristics