INDENG 242B final project, Spring 2026

Slime volleyball,
played by gradient descent.

We trained eighteen deep-RL policies on the same tiny volleyball court, then froze a 4,860-match benchmark to settle which one is actually best. Everything runs in your browser — watch the policies play, check the numbers, or pick up the keyboard and challenge them yourself.

18: trained policies
4,860: benchmark matches
0: servers involved

Preparing live inference…

— fps

— Loading…

5–5 LIVE · ON-DEVICE

— Loading…

Exhibition series Reference 0 · Trained 0 · Draws 0 Opening match

Policy Inspector per-action outputs · rolling 200 frames

Model controls & Policy Inspector

Left player Right player Simulation speed 2.00x

Inspect side Left Right

Output view Raw Centered

PPO exposes policy logits. DQN and Rainbow expose Q-values; “Centered” removes each frame’s mean. The RNN reference has neither and is not charted. Keyboard when a side is human — left: A/D move, W or Space jump · right: ←/→ move, ↑ jump · R restart.

Published once — this page can only read it

Official Rankings

Every pair of models played the same ten seeded matches, once per side. Rank comes from opponent-balanced Arena Score; the interval is a bootstrap over complete seed blocks, so a lucky serve can't move the table.

Verifying artifact…

Published: —
Protocol: —
Manifest hash: —
Completed: —

Official candidate rankings. Activate a model name for matchup details.
Rank	Model	Metric	95% CI	W–D–L	Vs ref	Life ±	Draw / timeout	Side Δ

Audit trail

Rankings can be fully recomputed from the raw match artifact. The schedule is fixed before play and does not react to results.

Download raw matches Read protocol Inspect model manifest

Your sandbox — results never leave this browser

Local Lab

Run your own seeded matches on your device and build a private rating table. Whatever happens here, the official rankings above don't move.

Local, not official

Model pool Featured (3) All models

Matchmaking Balanced Random

Starting seed

Browser-local exploratory ratings
#	Model	Type	Evidence	Rating	Games	Score	W–D–L	Vs ref	L/R

Slime volleyball,played by gradient descent.

Official Rankings

Audit trail

Local Lab

Slime volleyball,
played by gradient descent.