end-to-end-learning

This paper proposes RRDrive, a risk-aware occupancy representation for register-based end-to-end autonomous driving, which improves trajectory prediction and selection in complex scenes, achieving notable performance gains.

The paper proposes risk-aware occupancy as a dense representation for safety in end-to-end autonomous driving, introducing the ROIDrive network and RiskOcc4D-nuScenes dataset, which significantly reduces collision rates.

This paper presents gradCSCG, a differentiable reformulation of the Clone-Structured Causal Graph algorithm for end-to-end learning of cognitive maps directly from image sequences, combining it with a vector-quantized variational autoencoder.

A research paper proposing Color Pass-Through, an end-to-end learned framework that treats camera and display as a coupled system to improve color accuracy of images viewed on screens, achieving significant gains over baselines.

This paper introduces a concept-based interpretability framework for end-to-end autonomous driving models, using unsupervised dictionary learning via sparse autoencoders to decompose driving behavior into human-interpretable concepts. The framework enables analysis and targeted correction of model decisions, improving overall driving performance.

DiffSlack proposes a differentiable projection layer that enforces nonlinear inequality constraints in neural networks by reformulating them as equalities with learnable slack variables, achieving improved feasibility and planning success in vehicle path planning.

This paper proposes an end-to-end Conformer-based neural decoder for intracortical speech decoding from a participant with ALS, achieving a 23.80% character error rate without any external language model. It demonstrates that meaningful character-level decoding is possible in a fully end-to-end framework.

The authors introduce Sub-JEPA, a method using Subspace Gaussian Regularization to improve the stability of end-to-end world models like LeWM, showing consistent performance gains on continuous-control benchmarks.

MoCapAnything V2 introduces a fully end-to-end framework for arbitrary-skeleton motion capture from monocular video, jointly optimizing video-to-pose and pose-to-rotation predictions to resolve rotation ambiguity.