Shortcut Models Simplify Generative Planning for Offline RL

Researchers introduced Shortcut Trajectory Planning, or STP, for offline model-based reinforcement learning. Researchers propose Shortcut Trajectory Planning, a single-stage generative planner for offline reinforcement learning with an adjustable inference budget. The method is designed to reduce the repeated sampling cost associated with diffusion trajectory planners without requiring the separate teacher and student training used by consistency-based alternatives.

STP conditions a shortcut model on the requested step size. That lets one network generate trajectories with a smaller or larger inference budget rather than locking the planner to a fixed sampling procedure. Candidate state trajectories are ranked by a learned critic and feasibility penalty, then converted to actions through inverse dynamics. The feasibility term is intended to discourage plans that look valuable to the critic but are difficult to execute under the environment's constraints.

The training objective combines flow matching with recursive self-consistency across step sizes. In practical terms, the model learns both fine updates and larger jumps through the same training pipeline. This is the paper's main engineering distinction from consistency planners that first train a diffusion teacher and then distill a second model.

The authors evaluate STP across locomotion, navigation, manipulation, and dexterous-control tasks in the D4RL benchmark suite. They compare the method with actor-critic approaches and several generative planners. The manuscript reports competitive average results across the tested domains and presents ablations for warm-start planning and feasibility-aware selection. These results come from the research team and have not yet been independently reproduced.

The architecture offers a useful systems trade-off: a planner can adjust its generation budget at inference time and avoid a separate distillation pipeline. But the planning stack includes more than the generator. Candidate count, critic calibration, feasibility scoring, inverse-dynamics error, and environment latency can determine whether the complete loop is actually faster or safer.

LDS recommends measuring planning latency, control return, feasibility errors, and robustness beyond the authors' benchmark suite before production use. A strong follow-up would compare complete control-loop latency under an equal compute budget, test sensitivity to critic misranking, and evaluate behavior when dynamics differ from the offline data.