CogDrive: Cognition-driven multimodal prediction-planning fusion for safe autonomy
DOI: 10.26599/commtr.2026.9640016
archive: archived pipeline: cataloged verified
Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)
Summary
CogDrive addresses the challenge of safe autonomous driving in mixed traffic, where vehicles must navigate complex, multimodal interactions with human-driven agents. Existing methods face a dichotomy: learning-based approaches lack interpretability and struggle with rare, safety-critical behaviors, while rule-based systems lack adaptability in dynamic scenarios. To bridge this gap, the authors propose CogDrive, a cognition-driven framework that fuses multimodal prediction with safety-aware planning. The system aims to unify the adaptability of data-driven learning with the stability of rule-based reasoning, enabling interpretable decision-making under uncertainty. The framework consists of two coupled modules. The prediction module employs a cognition-driven approach using topological motion semantics to encode interaction modes (e.g., yielding, merging) based on relative angular displacement between agents. It utilizes an instance-centric coordinate system and a symmetric fusion encoder to capture bidirectional spatial dependencies among agents and map elements. A learnable query-based decoder, inspired by DETR architectures, generates multimodal trajectory hypotheses by associating queries with behavioral mode embeddings. The planning module adopts a safety-stabilized trajectory tree optimization. It first computes a short-horizon root trajectory to ensure immediate safety within replanning cycles, then extends mode-specific branches for long-term collision-free avoidance. This hierarchical strategy uses quadratic programming with dynamic safety constraints derived from the predicted interaction modes. Experiments on the Argoverse2 and INTERACTION datasets demonstrate that CogDrive achieves state-of-the-art performance. The model significantly reduces minimum Average Displacement Error (minADE) and miss rates compared to existing methods while maintaining trajectory smoothness. The use of a differentiable modal loss and max-margin classification objective effectively prevents mode collapse, allowing the network to learn sparse and unbalanced interaction behaviors. Closed-loop simulations further validate the system’s robustness, showing stable and adaptive behavior in strong-interaction scenarios such as merging and intersections. The framework successfully balances prediction accuracy with planning stability, avoiding the "freezing robot" phenomenon associated with overly conservative strategies. The significance of CogDrive lies in its establishment of an interpretable and reliable paradigm for safe autonomy. By coupling cognitive multimodal prediction with safety-oriented planning, the framework provides a coherent process for perceiving, anticipating, and acting under multimodal uncertainty. This approach overcomes the limitations of both pure learning-based and rule-based methods, offering a scalable solution for handling decentralized, multimodal interactions in complex traffic environments. The work highlights the potential of integrating explicit modal reasoning with optimization-based planning to enhance the safety and adaptability of autonomous vehicles.
Key finding
CogDrive achieves state-of-the-art performance in trajectory prediction accuracy and planning stability on benchmark datasets while maintaining safety in closed-loop simulations.
Methodology
simulation_modeling
Provenance
The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via author_sweep_intake on 2026-05-28.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | author_sweep | — | — | 2 | 2026-05-28 |
| archive | success | canonical_url | — | — | 1 | 2026-06-04 |
| extract | success | cached | — | — | 3 | 2026-06-10 |
| clean | success | clean | — | — | 1 | 2026-06-04 |
| chunk | success | chunk | — | — | 1 | 2026-06-04 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-06-04 |
| enrich | success | — | — | — | 1 | 2026-05-28 |
| promote | success | — | — | — | 1 | 2026-06-04 |
| summarize | success | llm | qwen3.6-27b-prismaquant | summ-v5 | 2 | 2026-06-10 |
| tag | success | vector_similarity | — | — | 15 | 2026-06-11 |
| verify | success | — | — | — | 2 | 2026-06-10 |
Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.
Topics
Ranked by relevance to this paper. Hover a topic for its definition.
Information type
What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).
- Theoretical Contribution: computational model