DEFENSIVE TACKLE
RAZE DT
Get through. Get to the quarterback.
Training study → Compare four methods, measured results & matched replays · 10M curriculum vs shaping → · Five-seed guard refinement → · New: unrestricted guard movement + fresh training →
THE DRILL
REP 001RUSH / DT0
5.00AWAITING SNAP
0PROTECT / G
DT INTENTRead the pocketChoosing a rushPosition: authored steering
G INTENTProtect the pocketChoosing a stancePosition: authored steering
LAST EVENT
Ready for the snap.
STATE MACHINES
DT / CURRENT STATEReady0.00s in state
G / CURRENT STATEReady0.00s in state
ENGAGEMENT AGE—
NEXT CHECK RESOLVES—
WINNER ODDS · G / DT—
Resolution risk grows with time. Winner odds apply only if the engagement resolves.
G BASE—
POCKET COMPRESSION—
SHOULDER LEVERAGE—
Base is modeled balance, not a win chance. Compression is guard displacement toward the QB since contact. Shoulder leverage measures the DT’s angle off the guard’s front.
Recent transitions · newest first
State machines control contact. Authored footwork keeps both players moving while engaged; learned opening tactics select the rush plan. Rewards remain +1 / −1.
QUICK / LIGHT
OFFENSIVE GUARD
ANCHOR G
Stay between the threat and the pocket.
EXPERIMENT DESK
Change a rule. Measure the difference.
Save a baseline, adjust player attributes or drill rules, then compare on identical seeded scenarios. Learning comparisons hold the rules fixed instead.
—Your first comparison will appear here.
WHERE IT CHANGES
| Scenario | Reps | Baseline | Current | Change |
|---|
CONTACT & MOVEMENT
| Diagnostic | Baseline | Current |
|---|
Intervals are 95% Wilson intervals for success rates in this simulation, not NFL predictions. A changed protection window changes the task: check the fraction of the window protected too. For a new check, change the seed. Better performance is not proof of more realistic technique.
Reports include the rules and frozen policies used. Data stays in this browser.
Experiment log · last 20 comparisons
| Hypothesis / change | Player | Reps / arm | Success | Run at |
|---|
IS PRACTICE HELPING?
Frozen reference tests after each training batch. DT and G face separate fixed opponents; their rates do not add to 100%. Lines break when settings change or learning resets.
Train a batch to start a learning history.
| Total reps | Checkpoint | DT test | G test | DT reward | G reward |
|---|
Tests use 180 matched scenarios per player. Reward columns are the mean opening-tactic reward earned during that practice batch, not win probabilities. Historic reps before this dashboard have no recorded checkpoints.
REWARD-BASED LEARNING
Try → score → update → repeat
Practice explores alternatives and updates two small tables. Watched reps and tests freeze those tables.
- Opening tactic · after the rep
- Both roles: +1 for a win, −1 for a loss.
- Positioning · during the rep
- No intermediate reward. Earlier choices learn from estimated future outcomes, including negative estimates.
- The last decision receives +1 for a win / −1 for a loss.
- What each player can read
- Relative position, starting tactic, time remaining, distance to the QB, and whether the players are moving freely or engaged.
Position values update at each decision, roughly every 0.3s while free movement is available, plus the terminal outcome. Learning rate 0.12; future-value discount 0.94. No distance, speed, survival-time, or block-count bonuses. Physical stats never grow from practice.
CURRENT EXPLORATION CHANCES
Practice continues from the saved tables. Recent reps can reverse old preferences; more practice does not guarantee improvement. The controller learns among authored moves.
LEARNED TACTICS & POSITIONING
Tactics are chosen at the snap; learned positioning adjusts targets during the rep. Signed tactic estimates run from −1 to +1; the center line is zero. These are learned reward values, not win probabilities.
RAZE / RUSH TACTICS
ANCHOR / PROTECTION
DT SPATIAL EXPERIENCE0 situations · 0 updates
G SPATIAL EXPERIENCE0 situations · 0 updates
RULES OF ENGAGEMENT
TUNABLEMovement → DT is Rushing; G is Blocking and aims to intercept the path to the QB. Both become Engaged when they meet within reach at a valid blocking angle.
Engaged → both keep moving: drive, anchor, shuffle, and shoulder attacks. Outcome checks run every 0.20s. The longer contact lasts, the more likely the next check is to resolve it.
Who wins → current odds if resolved: —. Each score uses 30% strength + 40% technique + 20% weight (scaled to 100 at 360 lb) + 10% speed. G chance = G score ÷ both scores. Equal attributes, including weight, give 50/50 odds. Equal rating averages alone do not: the contest weights strength more than speed.
Weight → adds mass ÷ 18 to the engagement score and 0.12 × mass to base contact power / resistance. It weights entry momentum and makes heavier players take less collision displacement. More weight lowers free-movement acceleration, but does not directly change top speed or the resolution clock.
Resolution pace → a smaller engagement timescale resolves contact sooner; a larger value allows longer contests. This is a probability curve, not a guaranteed duration. Check frequency changes when results can occur; it does not multiply the underlying risk.
After a result → DT wins: turn the corner. G wins: redirect the rush upfield. Both return to movement immediately. No flat-footed state or timed cooldown. A gap 0.25 yd beyond reach must open before a new engagement can start.
Contact lost → a range or angle loss releases both players without awarding a contest win. Pocket collapse can still reach the QB while engaged. These physical outcomes are separate from the stat roll.
Rep reward → only reaching the QB or protecting the full window earns +1 / −1. Winning an engagement adds no extra reward.
QB drop → 3.5–7.5 yd per rep, bell curve centered at 5.5 yd. The QB stays at that depth.
How the NPCs learn
This is reinforcement learning with fixed physical stats. A contextual bandit learns the opening tactic for each starting alignment. A tabular Q-learning controller then chooses a positional adjustment every 0.3 seconds during movement, using relative side, separation, depth, whether the tackle has passed the guard, the opening tactic, time remaining, distance to the QB, and contact availability. Explicit state machines determine when movement and contact are allowed: G Blocking ↔ Engaged; DT Rushing ↔ Engaged. Ready and Rep complete mark the boundaries of a rep.
Both layers use +1 for a win and −1 for a loss. Positioning has zero intermediate reward and learns from signed future-value estimates; no shaping bonuses are active. Both layers explore during practice and use learned preferences during watched reps and evaluation. Self-play or rule changes can improve or worsen performance.
This main lab uses outcome-only rewards. The separate training study tests shaping and three other methods. Saved reading: Ng, Harada & Russell (1999) · reward shaping PDF ↗.
The movement options and football rules are authored. The NPCs learn when to use them, including positional decisions during the rep. State transitions and contact probabilities are authored rules; the learned policy does not choose hand techniques yet. Contact movement is authored from coaching principles: pressure versus anchor, lateral footwork, and a body-separation constraint. Opening tactics affect the contact plan, but hand techniques and contact footwork are not independently learned. This is a 2D abstraction, without full-body physics. Ratings do not grow. There are no language models or external AI calls.
Coaching research & how it maps to this drill ↗
Changing attributes keeps your playbook but invalidates the benchmark. Reset learning for a clean experiment.
Build 014 lets the guard recover beside and behind the QB, with contact checks that follow the rush direction (engine 010). Default rating averages remain 77, with 70 technique each. Main-lab learning starts fresh for these rules; previous training remains available through “Export previous training.” Custom attributes are retained. The new study compares 15 fresh guards, with saved best checkpoints and replays.
Earlier training and experiment history remain saved separately. Use “Export previous training” below to keep a copy.
Training is local to this browser. Export a session to keep or share the experiment. Stylized football rules; no full contact-physics simulation.
Summary
We trained a guard and a tackle with simple reinforcement learning: each picks an opening tactic and adjusts its position, rewarded only for winning or losing the rep, with equal stats for both. Pre-trained policies are saved at every million reps up to 10M, and every score here is head-to-head on the same 500 test reps, so guard and tackle add up to 100%. What worked: the simulation, training speed (about twice as fast after removing slow copying), and saving reproducible checkpoints. What didn't: the tackle wins most reps at every stage, even untrained (guard 31%, tackle 69%), because reaching the QB within five seconds is easier than stopping every route; after 10M reps the guard only gets to about 38%, and results swing between checkpoints (guard 24% to 45%). A win/loss-only reward is too weak a signal for the guard. Next: a richer guard reward, and rules that make the matchup fairer.