Predictive but Not Plannable: RC-aux for Latent World Models

Abstract: A latent world model may achieve accurate short-horizon prediction while still inducing a latent space that is poorly aligned with planning. A key issue is spatiotemporal mismatch: these models are often trained with local predictive supervision, but deployed for long-horizon goal-directed search in latent spaces where Euclidean distance may not reflect what is reachable within a finite action budget. We present the Reachability-Correction auxiliary objective (RC-aux), a lightweight correction for this mismatch in reconstruction-free latent world models. RC-aux keeps the world-model backbone unchanged and adds planning-aligned supervision along two axes. Along the time axis, multi-horizon open-loop prediction trains the model beyond one-step consistency. Along the space axis, budget-conditioned reachability supervision, together with temporal hard negatives, encourages the latent space to distinguish states that are eventually reachable from those reachable within the current planning horizon. At test time, the learned reachability signal can also be used by a reachability-aware planner to favor trajectories that are both goal-directed and attainable under the available budget. We instantiate RC-aux on LeWorldModel and evaluate it under both continuation-training and matched-from-scratch settings. Across goal-conditioned pixel-control tasks and a LIBERO-Goal extension, RC-aux improves LeWM-style planning with modest additional cost. These results suggest that planning with latent world models depends not only on predictive accuracy, but also on whether the learned representation encodes the temporal and geometric structure required by downstream search. The code is available at https://github.com/Guang000/RC-aux.

• The paper introduces RC-aux, which improves latent world model planning by integrating multi-horizon open-loop prediction and budget-conditioned reachability supervision.
• It refines training by aligning predictive objectives with finite-horizon reachability, ensuring that the learned latent space supports feasible, goal-directed trajectories.
• Experiments across diverse environments, including LIBERO-Goal tasks, demonstrate enhanced planning performance and transferability in control tasks.

Summary of "Predictive but Not Plannable: RC-aux for Latent World Models"

Introduction

The paper "Predictive but Not Plannable: RC-aux for Latent World Models" (2605.07278) presents the Reachability-Correction auxiliary objective (RC-aux), designed to improve the planning capabilities of latent world models used in control from pixel-based observations. The main focus is on addressing the spatiotemporal mismatch existing between models trained for short-horizon predictions and their application in long-horizon planning scenarios. The authors highlight that accurate predictions alone are insufficient for planning; representation geometry that aligns with finite-horizon reachability is crucial.

Methodology

RC-aux Objective

RC-aux comprises a technique to modify latent world models without changing their backbone. The approach introduces planning-aligned supervision through multi-horizon open-loop prediction and budget-conditioned reachability supervision:

• Open-loop Prediction (Temporal Alignment): The model trains using multi-horizon predictions to overcome one-step consistency limitations, aligning the training with the planner's open-loop rollout needs.
• Reachability Supervision (Spatial Alignment): The latent space distinguishes reachable states within a finite horizon by employing budget-conditioned reachability predicates along with temporal hard negatives, which promote understanding of attainable vs. non-attainable states.

Integration with the Le WorldModel

RC-aux is implemented alongside Le WorldModel (LeWM) without altering its foundational design. This ensures that the model's latent dynamics are trained to represent planning-relevant geometry such as finite-horizon reachability.

Reachability-Aware Planning

During the planning phase, the learned reachability signal is employed, allowing the reachability-aware planner to prefer goal-directed trajectories that are feasible within the budget constraints. This mechanism shifts the focus from strict proximity to the goal to attaining it feasibly under the given action budget.

Experiments

The paper evaluates RC-aux across multiple environments: TwoRoom, Reacher, Push-T, Wall, and Cube, each characterized by goal-conditioned tasks that test the model's planning capacity. The RC-aux modifications lead to improved performance in four out of five scenarios compared to the LeWM family controls, with noteworthy enhancements on the Wall task, emphasizing the latent space's need to incorporate finite-reachability structures.

The study also extends the evaluation to the LIBERO-Goal tasks, involving a broader scope of robot manipulation scenarios, demonstrating that RC-aux benefits are transferable across different tasks beyond simple latent planning.

Conclusion

The authors conclude that RC-aux effectively enhances the latent world model's planning capability by aligning predictive objectives with planning requirements. This methodology not only focuses on predictive accuracy but also ensures that temporal and spatial geometries are tailored for goal-attainability. The future scope includes expanding reachability predicates into uncertainty-aware frameworks and richer directional distance measures. The paper underscores the significance of considering operational planning geometry in model training, thus paving the way for more reliable autonomous control systems in complex environments.