I Trained an AI to Beat Gorilla Tag’s Fastest Player

AI Training Setup & Goal

• The video describes training multiple AI-controlled gorillas in Gorilla Tag to discover the fastest possible movement strategy to beat an elite competitive pro player.
• Initial problem: the AI had no experience with walking—it initially moved by swinging arms extremely fast.

Reward / Learning Objective

To guide the AI, the presenter introduced reward signals and learning goals:

• Goal: reach the target balls.
• The AI was rewarded only when it reached the target faster than before.
• Method: reinforced learning via many repeated attempts, where the AI explores new hand/arm movement strategies and improves through trial and error.

Scale / Speed of Training

• 24 gorillas learning simultaneously.
• Training runs sped up at 20×.
• After a few hundred attempts, the AI’s movement improved substantially.

Making Movement More Human-Like (Product-Feature-Like Controls)

To reduce non-human motion, the presenter added two tuning parameters:

• Hand speed slider: limits arm speed.
• Hand smoothing slider: smooths motion to reduce jitter.

Effect:

• Early movement looked non-human.
• After additional tuning/attempts, the motion became more “legit” (less robotic/jittery).

Curriculum Learning: From Simple to Real Race Conditions

The AI’s objectives and training environment were expanded step-by-step.

Expanded Race Objective

• New goal: start on one side of the map and sprint to a ball as fast as possible.
• Reward: improved time each run.

Terrain Realism

• First training: flat surface (easier movement).
• Then training: bumpy/uneven terrain, where it initially got stuck on hilly sections.

Fix for Getting Stuck

• Adjusted the AI to move its hands up/down more.

Obstacle / Path Learning

• The presenter introduced hazard scenarios, including a pit from the real Gorilla Tag stump in forest.
• Then trained the gorillas to avoid objects that block their path on the Gorilla Tag Forest map (the map used by the competitive player).

Race Targeting & Performance Improvement

Route Matching

• The AI was assigned a race position matching the competitive opponent’s known path/route.

Benchmarking / Timeline

• Reported pro run time for that exact spot: ~9 seconds.
• AI behavior:
  • First attempt: roughly ~4 seconds (fast start)
  • Then it required longer learning (“grind”)
• After about an hour, time-lapses showed the route evolving from scattered movement to a near-straight optimal line.

Claimed Outcome

• The AI eventually produced a run that appeared to catch/approach the pro, then “beat” them.

Important Corrections: Bugs, Timing Logic, and Recalculations

Misidentified “Pro Player”

• The “pro player” the AI beat was later revealed to be the presenter (“me”), not the real pro, due to misunderstanding / previous setup.

External Reference (Undead) and Math Discrepancy

• The presenter cites community source Undead, who claims fastest achievable race time is ~2.35 seconds.
• A discrepancy was found: the reference time corresponded to a shorter race distance.

Distance / Time Comparison

• Undead’s ~2.35 s corresponded to 23.5 m.
• The presenter’s setup race length was 35.7 m (~52% longer).

The presenter recalculated:

• ( 23.5 \text{ m} / 2.35 \text{ s} \approx 10 \text{ m/s} )
• Applying 10 m/s to 35.7 m gives an expected time of about 3.67 s.

Conclusion: the AI was chasing an incorrect benchmark time.

Bug: Unfair “Flying” at the Start

• An early bug let gorillas fly at the start, giving an unfair speed advantage.
• After fixing the bug, the AI struggled to beat the target time.

Timing / Success Condition Fixes

• Success previously counted when within ~1.5 meters of the finish line; this was corrected.
• Additional issues causing flying were also fixed.

Final Behavior Insight (What the Optimal Run Looked Like)

After corrections and extended training, the AI produced an extremely fast run.

• The start reportedly looked different from the later motion:
  • At the very beginning, the AI used a “crazy” hand/arm move.
  • Then it transitioned into a more traditional competitive running style.
• The presenter notes that the pro-style segment was arguably one of the most optimal patterns the AI could converge on.

---

Main Speakers / Sources

• Main speaker: the video creator/presenter (the “I” who trained and tested the AI; later identified as “me” in the race).
• External source mentioned: Undead (video explaining fastest race time and providing the ~2.35 s reference with math).
• Other community reference: Horizon (person who pointed the presenter to Undead’s video).