Conditional Trajectory Peaks: Single-Pass Multimodal Policies over Action Chunks

2026-10-05

AI Summary

This study addresses the inherent conflict between execution diversity and replanning consistency in multimodal imitation learning by proposing the CTP framework. This framework pioneers single-trajectory modeling, enabling efficient policy generation through the joint prediction of action chunk candidates, probabilities, and scales. Furthermore, it introduces Distribution-Aware Peak Specialization (DAPS) and Evidential Gated Belief Transfer (ETBT) mechanisms to simultaneously preserve multimodal behavioral expressiveness and ensure closed-loop consistency. Experimental evaluations demonstrate that the proposed method achieves a 97.25% success rate on the LIBERO benchmark while reducing real-world robot inference latency to 75.8 ms, significantly outperforming existing baselines.

Abstract

Multimodal imitation learning requires diverse executable futures under the same observation and consistent behavior across replanning cycles. We present Conditional Trajectory Peaks (CTP), a single-pass policy framework that jointly predicts complete action-chunk candidates, probability masses, and trajectory scales. Distribution-Aware Peak Specialization (DAPS) specializes trajectory peaks using trajectory-level posterior responsibilities and mass- and scale-modulated overlap constraints. Evidence-Gated Trajectory Belief Transport (ETBT) maintains cross-chunk consistency through geometric correspondence between exchangeable candidate sets, while allowing current policy evidence to override historical constraints. CTP achieves a coverage score of 91.40% on Push-T; success rates of 100.0%, 79.72%, and 84.44% on D3IL Avoiding, Aligning, and Sorting-2, respectively. On LIBERO, CTP achieves an average success rate of 97.25%. In real-world dual-arm experiments, CTP preserves both placement modes in a two-plate task, succeeding in all 50 trials. On bottle uprighting and pen placement into a holder, it maintains success rates comparable to $\pi_{0.5}$ while reducing policy inference latency from 218.24 ms to 75.80 ms. These results demonstrate that single-pass trajectory modeling can combine multimodal behavior, closed-loop consistency, and efficient inference.

Problem

• Multimodal Imitation Learning
• Action Chunks
• Closed-loop Consistency
• Inference Latency

Innovation

• Conditional Trajectory Peaks
• Single-Pass Policy
• Multimodal Imitation Learning
• Action Chunks
• Distribution-Aware Peak Specialization