The authors propose Shortcut Trajectory Planning (STP), an offline model-based reinforcement learning framework that uses shortcut models to generate trajectories efficiently. This approach addresses the high inference costs of diffusion-based planners and the training instability associated with two-stage distillation in consistency-based methods.
- STP trains a conditional shortcut trajectory model in a single stage, avoiding complex multi-stage pipelines.
- The framework supports adjustable one-step and few-step inference through step-size conditioning.
- Candidate plans are selected using a critic augmented with feasibility-aware correction.
- Evaluations across standard D4RL benchmarks, including locomotion, navigation, manipulation, and dexterous control tasks, show strong performance.
STP simplifies the training pipeline for fast generative planning while maintaining competitive results on standard benchmarks.