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

Researchers present RC-aux, a lightweight auxiliary objective that improves latent world models for planning by addressing the spatiotemporal mismatch between short-horizon prediction training and long-horizon planning deployment. The method adds multi-horizon prediction and budget-conditioned reachability supervision to align learned representations with planning requirements, demonstrating improvements on goal-conditioned control tasks.

This research addresses a fundamental challenge in embodied AI: the gap between what makes a model good at prediction and what makes it useful for planning. Traditional latent world models optimize for one-step predictive accuracy, but this objective doesn't guarantee that the learned representation preserves the geometric and temporal structure necessary for effective planning. RC-aux tackles this through a dual-axis correction: temporal alignment via multi-horizon prediction ensures consistency across longer sequences, while spatial alignment through budget-conditioned reachability supervision teaches the model to distinguish states by their actual reachability rather than Euclidean distance.

The work builds on growing recognition within the robotics and reinforcement learning community that representation learning must be task-aligned. While predictive accuracy provides a useful training signal, it's insufficient for planning where action budgets constrain what states are actually achievable. By keeping the world-model backbone unchanged and adding auxiliary supervision, RC-aux maintains computational efficiency while improving downstream planning performance.

For the AI development community, this research provides practical guidance: achieving accurate predictions is necessary but not sufficient for planning-capable systems. The modest computational overhead combined with demonstrable improvements across multiple task domains suggests RC-aux represents a pragmatic advance in making latent world models deployable for real-world robotic control.

Developers working on embodied AI systems should monitor how this approach scales to more complex environments and whether reachability-aware planning becomes standard practice. The open-sourced implementation enables rapid adoption and extension by the research community.

• →Latent world models can predict accurately while remaining misaligned with planning requirements due to spatiotemporal mismatch
• →RC-aux adds multi-horizon prediction and budget-conditioned reachability supervision without modifying the core model architecture
• →The method teaches models to distinguish reachable states based on actual action budgets rather than geometric distance alone
• →Testing shows meaningful improvements on goal-conditioned pixel-control tasks with modest computational overhead
• →Reachability-aware planning at test time further improves trajectory quality by favoring attainable goal paths