Conditional Trajectory Peaks: Single-Pass Multimodal Policies over Action Chunks
CTP predicts multiple complete robot action futures in one pass while keeping replans consistent.
The paper frames the problem as imitation learning where the same observation can support several valid next trajectories. Its method predicts action-chunk candidates, their probability masses, and trajectory scales together, then uses DAPS and ETBT to specialize candidates and carry belief across chunks. Reported results include 97.25% average success on LIBERO and full success across 50 real-world dual-arm two-plate trials. The authors also report lower inference latency on bottle uprighting and pen placement, from 218.24 ms to 75.80 ms, while keeping success comparable to π₀.₅. HF Daily Papers' note
The paper frames the problem as imitation learning where the same observation can support several valid next trajectories. Its method predicts action-chunk candidates, their probability masses, and trajectory scales together, then uses DAPS and ETBT to specialize candidates and carry belief across chunks. Reported results include 97.25% average success on LIBERO and full success across 50 real-world dual-arm two-plate trials. The authors also report lower inference latency on bottle uprighting and pen placement, from 218.24 ms to 75.80 ms, while keeping success comparable to π₀.₅. HF Daily Papers' note
score 4