Fix: Godot NavigationAgent2D Stuck on Obstacle

A unit using NavigationAgent2D sets a target across the room, plots a path, and walks into the side of a crate that wandered into the corridor. It freezes there, body pressed against the obstacle, because every frame the agent says it has reached the next path point and tries to advance into geometry the path believes is empty.

Why Static Bakes Miss Dynamic Obstacles

When you bake a NavigationPolygon, Godot subtracts obstacle source geometry from the walkable region. Moving sprites and physics-driven bodies are not part of that snapshot. The bake produces a static polygon that knows the layout of walls and props that existed at bake time, nothing else.

The avoidance system is a separate runtime layer built on RVO2 (Reciprocal Velocity Obstacles). Each NavigationAgent2D contributes its position and velocity; each NavigationObstacle2D contributes a circular or polygonal exclusion zone. The agent’s requested velocity is filtered through this layer to produce a safe velocity that respects both static geometry (the polygon) and dynamic obstacles (other agents and obstacle nodes).

The Fix: Add NavigationObstacle2D Nodes

For each moving obstacle — a patrolling guard, a sliding crate, a destructible barrel — add a NavigationObstacle2D child:

# crate.gd
extends CharacterBody2D

func _ready():
    var obstacle = NavigationObstacle2D.new()
    obstacle.radius = 16.0   # half the crate’s width plus a 2px buffer
    obstacle.avoidance_enabled = true
    add_child(obstacle)

On the agent side, ensure avoidance is enabled and a callback handler updates the body velocity:

# enemy.gd
@onready var agent: NavigationAgent2D = $NavigationAgent2D

func _ready():
    agent.avoidance_enabled = true
    agent.radius = 14.0
    agent.velocity_computed.connect(_on_velocity_computed)

func _physics_process(delta):
    var next = agent.get_next_path_position()
    var desired = (next - global_position).normalized() * speed
    agent.set_velocity(desired)   # goes through avoidance

func _on_velocity_computed(safe_velocity):
    velocity = safe_velocity
    move_and_slide()

The critical detail: do not assign velocity = desired directly. Pass it to agent.set_velocity() and apply the value you receive in velocity_computed. That callback is where avoidance happens.

When the Geometry Itself Changes

If a wall is destroyed at runtime — not just moved — the static polygon needs a rebake. Don’t do this every frame; only after the change is committed:

func on_wall_destroyed(wall):
    wall.queue_free()
    await get_tree().process_frame   # let the node leave the tree
    $NavigationRegion2D.bake_navigation_polygon(true)   # on_thread=true

The on_thread=true argument moves the bake off the main thread. The agent continues using the previous polygon until the bake finishes and the new one is swapped in atomically.

Verifying

Enable Debug → Visible Navigation in the editor to see the walkable polygon highlighted in cyan. Run the game and watch agents reroute around obstacle nodes in real time. If an agent still walks into a crate, the crate likely lacks a NavigationObstacle2D child — or the obstacle’s radius is too small for the agent’s radius plus a margin.

Understanding the issue

This bug class falls into a pattern that's worth understanding beyond the specific case. In Godot Engine, the underlying behavior is shaped by how the engine layers its abstractions - the public API you call, the runtime systems that respond, and the platform-specific implementations underneath. A bug at any layer can produce symptoms that look like they originate at a different layer. Triaging effectively means recognizing which layer the symptom belongs to, even when the gameplay code is what's visible.

The specific bug described above is the kind that surfaces during integration rather than unit testing. It depends on a combination of factors: the asset configuration, the runtime state, the platform's specific behavior. In isolation, each piece looks correct; in combination, the bug emerges. This is why thorough integration testing - playing the actual game in realistic conditions - catches things that automated tests miss.

Why this happens

This bug class disproportionately affects late-stage development. The work to surface it is interactive testing in realistic conditions, which only really happens after the gameplay is in place and assets are populated. Catching it early requires deliberate testing of conditions that look unimportant.

At the engine level, the behavior comes from a deliberate design decision in Godot. The engine team chose a particular trade-off - usually performance versus convenience, or generality versus specificity - and that trade-off has consequences when you push against it. Understanding the trade-off is what turns 'this bug is mysterious' into 'this bug is the expected consequence of this design'.

Verifying the fix

After applying the fix, the verification step has three parts: confirm the original repro is resolved, confirm no obvious regressions in adjacent functionality, and (for shipping titles) deploy to a small player cohort first and watch the crash and report rates. Each step catches something the others miss.

Reproducibility is the prerequisite for verification. If you can't reliably reproduce the bug pre-fix, you can't reliably verify it post-fix. Spend time getting a clean reproduction before you write any fix code. The fix is fast once you understand the reproduction; the reproduction is the slow part.

Variations to watch for

There's almost always a less obvious case where the same problem applies. The reported case is the one a player hit; the related cases hide because they're rarer or affect fewer players. After fixing the reported case, search the codebase for the pattern - one fix often unlocks several.

Adjacent bugs often share a root cause. After fixing the case you've found, spend an hour searching the codebase for similar patterns. What's the same call with different arguments? The same data flow with a different entity type? The same lifecycle issue in a sibling system? Each match is a candidate for the same fix, or a related fix that prevents future bugs of the same class.

In production

In shipping builds, this issue may interact with other production-only behavior. Stripping, encryption, asset bundling, and platform-specific code paths can each modify the symptoms. When players report a related issue, capture build SHA, platform, and any feature flags - those three fields cover most of the production-only variations.

When triaging a similar issue in production, prioritize gathering data over hypothesizing causes. A player report describes a symptom; what you need is a build SHA, a session timestamp, and ideally a screen recording or session replay. With those, the bug becomes tractable. Without them, you're guessing at hypothetical reproductions that may not match what the player actually hit.

Performance considerations

If this issue manifests under high load (many actors, many particles, many network connections), profile the post-fix code path with realistic counts. The original cost was a bug; the new cost is real work, and real work has a budget.

Diagnostic approach

Before applying any fix, gather enough context to be confident you're addressing the actual cause and not a similar-looking symptom. The cheapest diagnostic step is reproducing the bug deterministically - if you can't get the same failure twice in a row, your fix attempts will be hard to evaluate. Lock down the reproduction first.

For Godot-specific diagnostics, the editor's profiler is the canonical starting point. Capture a representative frame with the symptom present; compare against a frame without the symptom; the diff often points directly at the cause. If the symptom is non-deterministic, capture multiple frames and look for the pattern - the cause is usually a state transition or a specific input value rather than a continuous effect.

Tooling and ecosystem

Modern engine versions ship better tooling for this kind of issue than older versions. If you're on an older release, the diagnostic step may take significantly longer because the tools you'd want don't exist yet. Sometimes the right answer is upgrading rather than fighting through limited tooling.

Within Godot, the relevant diagnostic surfaces include the standard frame debugger, memory profiler, and engine-specific debug overlays. Each one shows a different facet of what's happening. The frame debugger reveals draw call ordering and state transitions; the memory profiler shows allocation patterns; the debug overlay reveals per-system state. Bugs that resist one tool usually surrender to another - the trick is knowing which tool to reach for first.

Edge cases and pitfalls

Edge cases for this class of issue often involve specific timing: the first frame after a state change, the last frame before a transition, frames where multiple subsystems update simultaneously. Reproducing these reliably is part of what makes the bug class hard to test.

When writing a regression test for this fix, focus on the boundary conditions that surfaced the original bug. Tests that exercise the happy path catch obvious regressions; tests that exercise the boundary catch the subtler regressions that look like new bugs but are really the original returning. The latter are the tests that earn their keep over the long life of the project.

Team communication

Document the fix and its rationale in the commit message or attached engineering doc. Future engineers will encounter related issues; the rationale tells them whether your fix is reusable or specific to the case at hand. Without rationale, the fix gets reverted or copied incorrectly.

If this fix touches a system several engineers work in, a short writeup in the team's engineering channel helps. Not a full design doc - a paragraph explaining what was wrong, what's fixed, and what to watch for. Future engineers encountering similar symptoms will search for the fix; making it findable is a small investment that pays back later.

“Static polygon for walls, obstacle nodes for movers, avoidance callback for velocity. Three layers, each with one job.”

Always wire velocity_computed — bypassing it means avoidance never runs.