Previous Work
Phase 1 of this research focused on integrating I-JEPA self-supervised features with lightweight MLP planning heads, achieving 82.06% PDMS using only camera inputs and demonstrating that simple architectures with strong SSL features can match complex multi-modal systems.
About & Methodology
Research details, architecture diagrams, key findings, and engineering highlights from the I-JEPA + MLP planning study.
View →Experiment Results
40+ experiments comparing I-JEPA, TransFuser, DINO, and DINOv2 backbones with detailed performance analysis.
View →Interactive Demo
deck.gl visualization of real-time trajectory planning with model selection and performance metrics.
View →