Conditional Trajectory Peaks for Single-Pass Multimodal Imitation Learning
October 4, 2026
The Conditional Trajectory Peaks (CTP) framework uses Distribution-Aware Peak Specialization (DAPS) to predict complete action-chunk candidates in a single pass. CTP achieved a 91.40% coverage score on the Push-T benchmark while maintaining cross-chunk consistency via Evidence-Gated Trajectory Belief Transport.
HOW THIS AFFECTS YOU
●
builderYou can implement more consistent multimodal imitation learning policies that handle diverse executable futures.
●
researcherYou can study the effects of trajectory-level posterior responsibilities on imitation learning stability.