Refactor waypoint structure and metrics

[vcharraut]

Summary

• Move evaluation metric accumulators from Agent state into per-agent logs.
• Replace the old ttc_tfl metric slot with distance_to_collision and update visualization labels.
• Remove unused waypoint curvature/TTC cache state while computing TTC metrics directly from metric arrays.
• Keep puffer score display fields updated during the episode.

[Copilot]

Pull request overview

Refactors waypoint/metrics plumbing in the Drive environment by removing waypoint curvature (kappa), moving some per-episode evaluation accumulators from Agent into Log, and rewiring TTC-related metrics to expose distance_to_collision instead of the prior ttc_tfl metric.

Changes:

• Replace ttc_tfl metric wiring with distance_to_collision across C metrics and Python viz labels.
• Remove Waypoint.kappa and related curvature computation/export from the C waypoint API.
• Shift puffer/eval accumulator updates (wrong-way distance, speed violations, multi-lane time, TTC sampling) to Log and plumb log_idx through metric computation.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
pufferlib/viz.py	Updates the metric label list to use distance_to_collision in place of ttc_tfl.
pufferlib/ocean/drive/drive.h	Refactors TTC + eval metric accumulation and removes waypoint curvature; updates offroad neighbor offsets constants and metric/reward plumbing.
pufferlib/ocean/drive/datatypes.h	Updates metric indices (adds DISTANCE_TO_COLLISION_IDX), removes Waypoint.kappa, and relocates traffic_control_in_scope.
pufferlib/ocean/drive/binding.c	Stops exporting the removed waypoint field kappa to Python.

Comments suppressed due to low confidence (2)

pufferlib/ocean/drive/drive.h:593

• compute_log_yaw_rate divides by dt but no longer guards against dt <= 0, which can lead to division-by-zero / inf yaw rates if dt is misconfigured (e.g., from INI). Please restore the defensive check (and keep the parameter const since the agent isn't mutated here).

    int has_next = (next_t < agent->trajectory_length) && (agent->log_valid[next_t] == 1);

    if (has_prev && has_next) {
        float dtheta = normalize_heading(agent->log_heading[next_t] - agent->log_heading[prev_t]);
        return dtheta / (2.0f * dt);

pufferlib/ocean/drive/drive.h:1836

• compute_new_route removed clamping of env->num_target_waypoints before using it to derive min_route_distance. Since num_target_waypoints can come from INI parsing (see drive.c) without bounds checks, this can produce extreme route distances and unstable routing behavior. Please clamp (and ensure a positive value) locally like other goal/obs code paths do.

    int num_target_waypoints = env->num_target_waypoints;
    float min_route_distance;
    // NOTE: make both multipliers config values and tune from a metric (route regenerations per 1k env steps).
    if (env->target_type == TARGET_STATIC) {
        min_route_distance = env->max_waypoint_spacing * num_target_waypoints * 2.0f;

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[WaelDLZ]

It's only refactoring I trust you

[eugenevinitsky]

think collision_offsets was maybe clearer?

[vcharraut]

I disagree with that since collision is meant for objects also

[eugenevinitsky]

think the more natural version is agent_logs; log_agent implies an agent that is used for logging