Shortcut Trajectory Planning Boosts Offline RL Efficiency

Diffusion-based trajectory planners have shown impressive results in offline reinforcement learning (RL), but their iterative denoising process often leads to high inference costs. While consistency-based planners reduce sampling steps, they typically rely on a two-stage teacher-student distillation, which complicates training and can introduce instability. This paper introduces Shortcut Trajectory Planning (STP), an offline model-based reinforcement learning framework designed for efficiency. STP trains a conditional shortcut trajectory model in a single stage, streamlining the training process. A key feature of STP is its support for adjustable one-step and few-step inference, achieved through step-size conditioning. It also enhances plan selection with a critic augmented by feasibility-aware correction. Evaluated across standard D4RL benchmarks, including locomotion, navigation, manipulation, and dexterous control tasks, STP demonstrates strong performance while simplifying the generative planning pipeline for faster inference.