Research map / Probabilistic and scientific modeling / Dynamical systems

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

NeurIPS 2026 · Probabilistic and scientific modeling › Dynamical systems · approach: Latent-space control and planning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RC-aux, a backbone-agnostic auxiliary objective for LeWorldModel that adds multi-horizon open-loop prediction and budget-conditioned reachability supervision (with temporal hard negatives), plus a reachability-aware planner, improving goal-conditioned pixel control on several tasks.

Problem

Latent world models trained with local one-step predictive supervision can yield latent geometries (Euclidean distance) that do not reflect finite-budget reachability, hurting long-horizon goal-directed planning from pixels.

Approach

Replace one-step latent prediction with multi-horizon open-loop prediction loss and add a budget-conditioned reachability predicate trained on trajectory-offset labels, using temporal hard negatives (eventually reachable but outside budget) to make the budget identifiable; at test time, discount the latent planning cost by a reachability score.

Key results

• RC-aux reaches 83.6% success on Wall vs 50.4% for LeWM (from-scratch), the largest gain.
• RC-aux improves TwoRoom to 98.0% vs 88.8% for LeWM-cont, and Cube to 76.0% vs 72.8%; improves four of five matched comparisons.
• On LIBERO-Goal with an OFT-style head, RC-aux raises mean success from 0.712 (LeWM) to 0.812, and to 0.864 with repeat tuning.

Surprising findings

• RC-aux with lambda_plan=0 (training-only effect) already lifts Wall from 50.4% to 72.4%, indicating the training objective alone drives much of the gain.
• Push-T shows no improvement (90.8% vs 91.2% for LeWM-cont), so the method does not help uniformly.

Limitations

• Reachability labels are trajectory-induced proxies rather than true environment reachability; shorter unobserved paths may exist.
• Planner uses a simple multiplicative reachability gate rather than a principled feasibility treatment.
• Results rely on local runs with five evaluation groups and limited benchmarks; Push-T shows no gain and LIBERO results use an OFT-style head rather than the planner.

Experimental setup

Datasets: TwoRoom Reacher Push-T Wall Cube LIBERO-Goal

Benchmarks: TwoRoom Reacher Push-T Wall Cube LIBERO-Goal

Metrics: success rate mean success ms per planner cost call

Baselines: LeWM LeWM-cont DINO-WM PLDM GCBC IQL IVL DINO-WM+prop OpenVLA-OFT

Builds on: LeWorldModel DINO-WM PLDM SIGReg OFT

Closest papers by meaning

• Sparse Planning in Visual World Models via Cost Gradients (NeurIPS 2026 · similarity 0.89)
• Coarse-to-Fine Compositional Diffusion for Long-Horizon Planning (NeurIPS 2026 · similarity 0.87)
• Simplifying Latent Dynamics with Softly State-Invariant World Models (NeurIPS 2024 · similarity 0.87)
• HaM-World: Soft-Hamiltonian World Models with Selective Memory for Planning (NeurIPS 2026 · similarity 0.87)
• PcLast: Discovering Plannable Continuous Latent States (ICML 2024 · similarity 0.87)
• Temporal Difference Flows (ICML 2025 · similarity 0.86)