Hierarchical Planning Boosts Long-Horizon Robot Control by Up to 14.7 Points

A new robotics paper accepted at WM@Booth 2026 shows that freezing a pretrained world model and layering a planner on top of it can meaningfully close the gap in long-horizon manipulation tasks. The method, called Hi-LeWM, extends an existing latent world model known as LeWorldModel (LeWM) by introducing a high-level component that plans over latent subgoals while leaving the underlying low-level model untouched.

The core idea addresses a familiar weakness in goal-conditioned control: flat world models tend to perform well when the target state is close to the current one, but their accuracy degrades sharply as the distance between start and goal grows. Hi-LeWM tackles this by decomposing a distant goal into a sequence of intermediate latent subgoals, letting the frozen low-level LeWM handle short segments of the trajectory while a separate high-level module decides where those segments should lead.

What the Results Show

The researchers tested Hi-LeWM on two manipulation benchmarks, PushT and Cube, across a range of increasing goal offsets designed to simulate progressively harder long-horizon scenarios. The gains scale with task difficulty in a telling way. At medium-range horizons, Hi-LeWM outperformed the flat LeWM baseline by 11.3 percentage points. At the longest offsets tested on PushT, the improvement widened to 14.7 percentage points, suggesting the hierarchical approach becomes more valuable precisely as tasks get harder, which is where flat models typically struggle most.

Interestingly, the paper also finds that hierarchy isn't always the answer. For short-horizon goals, the best-performing Hi-LeWM configuration collapses its high-level planning to a single step, effectively behaving close to the flat baseline. This detail matters for practitioners: it implies the high-level horizon should be treated as a tunable parameter rather than fixed, since over-planning on easy tasks can introduce unnecessary complexity without a corresponding performance payoff.

Why the Architecture Choice Matters

Freezing the pretrained low-level model rather than retraining it end-to-end is a deliberate design decision. It lets the team isolate the contribution of the high-level planner, avoid the instability that often comes with jointly optimizing multi-level systems, and reuse a model that was presumably expensive to train in the first place. This mirrors a broader trend in robot learning where practitioners try to get more mileage out of existing world models through modular add-ons rather than retraining from scratch.

The paper is filed under Robotics, Artificial Intelligence, and Machine Learning categories on arXiv, reflecting its dual relevance to both the world-model research community and applied robotic manipulation. While PushT and Cube are simulated benchmark tasks rather than real-world deployments, the consistent scaling of improvement with horizon length gives a concrete signal that hierarchical latent planning is a viable lever for extending the reach of learned world models beyond their native short-horizon competence.