Out-of-bounds & obstacle collision

track_gen.collision answers two questions every batched sim asks: did the agent leave the drivable band? and did it hit a point obstacle?

Out-of-bounds checking

track_gen.collision.CollisionChecker answers, per oriented box, whether it left the drivable band — with signed clearance, nearest boundary point, and inward normal. Two backends: the exact segments scan (default; no precompute, reads regenerated tracks automatically) and baked sdf grids (O(1) queries after a per-regeneration bake(); distances accurate to about one grid cell). See Performance below for measured numbers; the rule of thumb: segments unless a track batch serves hundreds of queries between regenerations.

The full per-box contact report: OOB flag (color), signed clearance (label), nearest boundary point (dotted) and inward normal (arrow).

Performance

Measured with benchmarks/benchmark_collision.py on an RTX 4090 (warp-lang 1.14, E = 8192 generated tracks × 8 boxes = 65,536 box queries per call, default TrackGenConfig):

backend	eager / query	graph replay / query	precompute per regeneration
segments (exact)	~0.32 ms	~0.29 ms	none
sdf (128² grids)	~0.053 ms	~0.025 ms	~49 ms bake() + ~670 MB

Accuracy of sdf against the exact backend on the same workload: the signed-distance error is typically ~2% of a grid cell (p99 < 0.4 cells); OOB flags agree on 99.6% of boxes, disagreeing only inside the ±1-cell band around zero clearance. SDF normals are noisy near the middle of the band (medial axis) and at sharp features — prefer segments when normals or nearest points drive a contact response.

Rule of thumb: segments is the default — exact, no memory, no rebake, and already ~5 ns per box. Switch to sdf when a track batch serves more than roughly 200 queries between regenerations and only the OOB flag and approximate clearance are consumed (e.g. RL reward shaping).

Reproduce:

python -m benchmarks.benchmark_collision --E 8192

Disc obstacles (gate posts, cones)

track_gen.collision.DiscChecker checks oriented boxes against disc obstacles. Gate posts are two lines of code:

import numpy as np, warp as wp
from track_gen.collision import DiscChecker

dev = seq.left.device                       # bind posts on the same device
posts = np.empty((E, 2 * G, 3), np.float32)
posts[:, 0::2] = seq.left.numpy().reshape(E, G, 3)
posts[:, 1::2] = seq.right.numpy().reshape(E, G, 3)
checker = DiscChecker(wp.array(posts.reshape(-1, 3), dtype=wp.vec3f,
                               device=dev),
                      radius=0.03, max_boxes=1, num_envs=E)

NaN padding in the gate arrays carries over and NaN discs are skipped, so no per-env count bookkeeping is needed. A hit reports the deepest disc; for interleaved posts, gate = disc // 2.

posts is a host-side snapshot taken via .numpy() — it does NOT alias seq.left/seq.right, so a later regeneration is invisible to it; rebuild posts (repeat this recipe) after every generate(). The Course facade’s built-in post_radius option does this rebuild device-side, automatically, on every generate().

Gate posts as discs; the same checker makes cones physical.

API: track_gen.collision.CollisionChecker, track_gen.collision.DiscChecker — see the API reference.