McMaster University
Planning-Aware Trajectory Prediction: Routing Downstream Planning Risk into a Frozen-Perception Multimodal Predictor
Abstract
dc:description.abstractAutonomous vehicles forecast how surrounding agents will move and then plan a safe trajectory, yet the two stages are optimized for different things. Trajectory predictors are trained and ranked by displacement error, whereas a planner is served by forecasts that help it avoid collisions and act smoothly. Displacement is therefore a proxy for prediction quality rather than a measure of planning value, and the gap between them is usually left for the planner to absorb. This thesis introduces the Planning-Aware Loss, a plug-in, differentiable training signal that routes a downstream planning-risk gradient into the mode distribution of a multimodal trajectory predictor. The central question is whether such a signal reshapes the predictor or merely regularizes the planner that consumes its forecasts. To answer it cleanly, the loss is compared against a stop-gradient control that receives the identical objective with only the gradient path to the predictor severed, under a controlled, frozen-perception protocol on the nuScenes dataset spanning oracle and perceived input domains. The full gradient path separates from the control on all four conditions, with non-overlapping bootstrap 95% confidence intervals for a collision proxy at a displacement-controlled working point, and the separation holds across two prediction carriers, three differentiable risk surrogates, and both domains. Under a prediction-sensitive planner trained once and then frozen, open-loop displacement turns in favour of the full loss, consistent with a ghost-braking-like effect. The result is a controlled demonstration rather than a leaderboard claim, bounded to open-loop evaluation, frozen perception, and a collision-only risk term, and motivated throughout by safety at complex urban intersections.
Degree
thesis:*- Department dc:contributor.department
- Electrical and Computer Engineering
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Guo, Xiyuan
- Advisor dc:contributor.advisor
-
- Emadi, Ali
Subjects
dc:subject × 5Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Identifier URI
- https://doi.org/10.71548/2267