RTS Pathfinding Explained (and Why It Feels So Good When It Works)

Storyline

• No game storyline is directly discussed.
• The video focuses on the RTS feel of crowd simulation and pathfinding, not narrative plot.

Core topic: Pathfinding vs. Crowd Simulation (what “feels good”)

Pathfinding (RTS meaning)

• Primarily routes around static obstacles (terrain/buildings) using:
  • nav meshes
  • path queries

Why pathfinding doesn’t cover unit-vs-unit navigation

• In RTS, units generally don’t path around other units.
• Updating nav meshes for moving units is too expensive, especially with:
  • higher unit speeds
  • larger unit counts

Crowd simulation (the missing piece)

• Crowd simulation handles local avoidance using limited nearby information—often framed as something like:
  • “the 3 ft in front of you” concept
• Good crowd behavior prevents:
  • frustrating jams and deadlocks
  • walling/clumping
  • “give up” cases where units stop trying to reach goals

Gameplay / behavioral highlights (examples from RTS games)

StarCraft II–style behavior (described as strong)

• Units dynamically reposition to reduce interference.
  • Example: lower-range units avoid pushing into higher-range units.
• Units can “go around” enemies rather than only reacting in place.
• Caveat: SC2 still has edge cases where units become blocked and abandon attacking because they can’t find/“see” alternate routing.

Age of Empires 4–style approach (described as different)

• Uses a pushing mechanic upon entering attack range.
• Side effect: front units can become too close.
  • Example mentioned: archers pushed near catapults leading to full impact.
• Mitigation:
  • Formations / pre-sorting arrange units by role (e.g., knights up front, archers behind).
  • This reduces problematic pushing and lessens reliance on complex avoidance.

Key issues / edge cases discussed

• Units stuck behind “invisible non-navmesh blockers”
  • Concept example: marines may behave as if routing is impossible because local crowd logic limits what they can infer.

• Units “block” each other

  • These blockers may not behave like true static obstacles for navmesh logic.

• Traffic jams at formations and choke points

  • Straight walls or concave arrangements can cause groups to “second guess” movement due to limited local awareness.
  • At choke points with multiple routes, units may overuse a single lane until they believe they’re stuck.
  • The video claims a significant portion can give up.

• Unwanted movement of units meant to stay put

  • Units placed for defense/positioning can be displaced by passing units.
  • Example: hydra/movers causing an ultralisk-area unit to drift away from its intended spot.

• Small-unit vs. big-unit clumping

  • Without strong avoidance, small units can get caught against large units.
  • This is especially likely if large units form a wall.

• RVO avoidance limitations (reciprocal velocity obstacle)

  • RVO may work for small numbers.
  • At massive unit counts it breaks down:
    • units can be left behind
    • units may fail to reach destinations

Strategies / techniques mentioned to improve crowd simulation (including Zerospace)

Collision tiers (Zerospace)

• Prevent larger units from being displaced by smaller units around them.

RVO-style avoidance (not sufficient alone at scale)

• Redirects velocities as units approach each other.
• But performance/behavior degrades with large crowds.

Vehicles / acceleration-based units

• Using acceleration changes motion behavior and enables different tactical outcomes.
  • Example: hover tanks.

Precision and control features

• Units move toward the center of the mouse click location.
• “Magic box implementation” to preserve intent when clumps are small enough.
• “One pixel away” adjustments for cases where targets require that precision.

Formations (planned / used)

• Reduce clumping using role-based grouping behaviors.
• Intended to create controllable large-scale visual patterns (e.g., “Roman legion look”).

Performance / tick-rate considerations (why it’s hard)

• Crowd simulation is expensive:
  • computational cost rises sharply with more units
• “Limited vision” behavior is partly a performance necessity.

Mentioned performance examples / targets

• StarCraft II
  • Drops below 30 FPS around ~1,000 marines attacking
  • Tick rate ~23 per second mentioned
• Zerospace
  • Handles ~500 units at ~80 tick rate on a described 2.4 GHz machine
  • Tick rate may be lowered later (40–60 mentioned)

Optimization tactics mentioned (H4-style context)

• Lower crowd-sim tick frequency:
  • rotations/direction changes happen less often
  • transitions remain smoother
• Use one CPU for the crowd sim step to reduce hardware requirements.

What better crowd simulation unlocks (game design impact)

• More RTS spectacle:
  • more units swarming and “smashing” feels better
• Especially benefits PvE modes:
  • campaigns
  • co-op scenarios
• Zerospace survival-mode example
  • Similar vibe to They Are Billions waves, but co-op
  • Improved crowd sim would allow thousands of units in wave nights.

Sources / games mentioned

• Marv (from Zerospace / Zerpac(e))
• StarCraft / StarCraft 2
• Age of Empires 4
• Warcraft 3
• They Are Billions
• Zerospace / Zerpac (featured game discussed)