Research map / Reinforcement learning / Visual and representation RL

SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning

NeurIPS 2026 · Reinforcement learning › Visual and representation RL · approach: Graph search and compositional goals · 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 SAGAS, which combines a frozen GAS-style offline latent reachability graph and executor with task-time semantic graph augmentation and cost-aware Büchi product search to zero-shot plan and execute LTL tasks from fragmented offline data.

Problem

Executing LTL-specified long-horizon robotic tasks in a strictly offline, model-free setting using only fixed task-agnostic trajectory fragments, without dynamics models, task demonstrations, or online interaction.

Approach

Repurposes graph-assisted stitching (GAS) as a reusable latent reachability substrate, lifts the frozen graph with test-time proposition semantics (soft labels from raw-state support and anchor witness nodes), and performs A*-style Büchi product search for accepting prefix-suffix waypoint plans executed by a frozen goal-conditioned policy.

Key results

• SAGAS achieves 75.4% SR on AntMaze Easy and 65.0% on Hard, versus 55.9% for LF-GAS on Hard AntMaze.
• On HumanoidMaze Hard, SAGAS reaches 49.5% SR vs 43.2% for LF-GAS and 37.7% for AP-HIQL.
• SAGAS attains the lowest NCC in every domain-difficulty group (e.g., 0.38 vs 0.46 for LF-GAS on AntMaze Hard).

Surprising findings

• Only 6.2% of SAGAS failures stem from absence of a feasible product prefix or suffix; most failures come from low-level execution issues such as stalling or overturning.

Limitations

• Success is bounded by support of the fixed offline dataset; propositions without dataset witnesses are unavailable.
• Soft-label negative-literal screening is node-level and does not certify intermediate states avoid forbidden regions.
• Low-level frozen executor realization over long horizons is the main remaining failure source.

Experimental setup

Datasets: OGBench antmaze navigate OGBench antmaze stitch OGBench antmaze explore OGBench humanoidmaze navigate OGBench humanoidmaze stitch

Benchmarks: OGBench antmaze OGBench humanoidmaze (LTL task suites)

Metrics: finite-lasso execution success rate (SR) normalized capped cost (NCC)

Baselines: LF-GAS GA-GAS AP-HIQL

Builds on: GAS HIQL DDPG+BC IQL ltl2ba

Closest papers by meaning

• TeLoGraF: Temporal Logic Planning via Graph-encoded Flow Matching (ICML 2025 · similarity 0.90)
• Embodied Task Planning via Graph-Informed Action Generation with Large Language Models (ICML 2026 · similarity 0.89)
• Zero-Shot Trajectory Planning for Signal Temporal Logic Tasks (NeurIPS 2025 · similarity 0.89)
• Plan-Seq-Learn: Language Model Guided RL for Solving Long Horizon Robotics Tasks (ICLR 2024 · similarity 0.88)
• Graph-Assisted Stitching for Offline Hierarchical Reinforcement Learning (ICML 2025 · similarity 0.88)
• Hierarchical Planning for Complex Tasks with Knowledge Graph-RAG and Symbolic Verification (ICML 2025 · similarity 0.88)