Rectify, Don't Regret: Avoiding Pitfalls of Differentiable Simulation in Trajectory Prediction

Current open-loop trajectory models struggle in real-world autonomous driving because minor initial deviations often cascade into compounding errors, pushing the agent into out-of-distribution states. While fully differentiable closed-loop simulators attempt to address this, they suffer from shortcut learning: the loss gradients flow backward through induced state inputs, inadvertently leaking future ground truth information directly into the model's own previous predictions. The model exploits these signals to artificially avoid drift, non-causally "regretting" past mistakes rather than learning genuinely reactive recovery. To address this, we introduce a detached receding horizon rollout. By explicitly severing the computation graph between simulation steps, the model learns genuine recovery behaviors from drifted states, forcing it to "rectify" mistakes rather than non-causally optimizing past predictions. Extensive evaluations on the nuScenes and DeepScenario datasets show our approach yields more robust recovery strategies, reducing target collisions by up to 33.24% compared to fully differentiable closed-loop training at high replanning frequencies. Furthermore, compared to standard open-loop baselines, our non-differentiable framework decreases collisions by up to 27.74% in dense environments while simultaneously improving multi-modal prediction diversity and lane alignment.

Task Weighting through Gradient Projection for Multitask Learning

In multitask learning, conflicts between task gradients are a frequent issue degrading a model's training performance. This is commonly addressed by using the Gradient Projection algorithm PCGrad that often leads to faster convergence and improved performance metrics. In this work, we present a method to adapt this algorithm to simultaneously also perform task prioritization. Our approach differs from traditional task weighting performed by scaling task losses in that our weighting scheme applies only in cases where tasks are in conflict, but lets the training proceed unhindered otherwise. We replace task weighting factors by a probability distribution that determines which task gradients get projected in conflict cases. Our experiments on the nuScenes, CIFAR-100, and CelebA datasets confirm that our approach is a practical method for task weighting. Paired with multiple different task weighting schemes, we observe a significant improvement in the performance metrics of most tasks compared to Gradient Projection with uniform projection probabilities.