{"paper":{"title":"FutureSightDrive: Thinking Visually with Spatio-Temporal CoT for Autonomous Driving","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Generating one future visual frame lets driving models plan trajectories by preserving spatial and temporal details that text chains of thought discard.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Mengwei Xie, Mu Xu, Ning Guo, Shuang Zeng, Xing Wei, Xinran Liu, Xinyuan Chang, Yifan Bai, Zheng Pan","submitted_at":"2025-05-23T09:55:32Z","abstract_excerpt":"Vision-Language-Action (VLA) models offer significant potential for end-to-end driving, yet their reasoning is often constrained by textual Chains-of-Thought (CoT). This symbolic compression of visual information creates a modality gap between perception and planning by blurring spatio-temporal relations and discarding fine-grained cues. We introduce FSDrive, a framework that empowers VLAs to \"think visually\" using a novel visual spatio-temporal CoT. FSDrive first operates as a world model, generating a unified future frame that combines a predicted background with explicit, physically-plausib"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"This imagined scene serves as the visual spatio-temporal CoT, capturing both spatial structure and temporal evolution in a single representation. ... our visual spatio-temporal CoT bridges the perception-planning gap, enabling safer, more anticipatory autonomous driving.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The generated future frame is physically plausible and contains the exact spatio-temporal cues needed for accurate inverse-dynamics planning; if the predicted lanes or boxes are systematically wrong, the planning step will inherit those errors.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"FSDrive uses a generated future scene frame as visual spatio-temporal CoT to improve VLA models for safer autonomous driving trajectory prediction.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Generating one future visual frame lets driving models plan trajectories by preserving spatial and temporal details that text chains of thought discard.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"35cedf2a080e960e3b549915a844e95069001ed4bd56d519a80ce86bb07f9266"},"source":{"id":"2505.17685","kind":"arxiv","version":3},"verdict":{"id":"65212241-e689-4a70-9f13-89496a316b1a","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T19:15:54.700395Z","strongest_claim":"This imagined scene serves as the visual spatio-temporal CoT, capturing both spatial structure and temporal evolution in a single representation. ... our visual spatio-temporal CoT bridges the perception-planning gap, enabling safer, more anticipatory autonomous driving.","one_line_summary":"FSDrive uses a generated future scene frame as visual spatio-temporal CoT to improve VLA models for safer autonomous driving trajectory prediction.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The generated future frame is physically plausible and contains the exact spatio-temporal cues needed for accurate inverse-dynamics planning; if the predicted lanes or boxes are systematically wrong, the planning step will inherit those errors.","pith_extraction_headline":"Generating one future visual frame lets driving models plan trajectories by preserving spatial and temporal details that text chains of thought discard."},"references":{"count":98,"sample":[{"doi":"","year":2020,"title":"H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Krishnan, Y . Pan, G. Baldan, and O. Beijbom. nuscenes: A multimodal dataset for autonomous driving.CVPR, 2020","work_id":"15713acc-d349-49db-91d1-7d2900bda05d","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"X. Chang, M. Xue, X. Liu, Z. Pan, and X. Wei. Driving by the rules: A benchmark for integrating traffic sign regulations into vectorized hd map.CVPR, 2025","work_id":"46d82771-2c7b-4644-9cae-89dcb58e2909","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"VADv2: End-to-End Vectorized Autonomous Driving via Probabilistic Planning","work_id":"e7670f83-e1e1-41e7-86eb-39477a3a10b2","ref_index":3,"cited_arxiv_id":"2402.13243","is_internal_anchor":true},{"doi":"","year":2025,"title":"Y . Chen and R. Greer. Technical report for argoverse2 scenario mining challenges on iterative error correction and spatially-aware prompting.arXiv preprint arXiv:2506.11124, 2025","work_id":"264fb541-de37-4b71-869c-d909712f1cd0","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Y . Chen, Y .-Q. Wang, and Z. Zhang. Drivinggpt: Unifying driving world modeling and planning with multi-modal autoregressive transformers.ICCV, 2025","work_id":"4fee8c4d-b4e8-4356-a454-8ce64833a17b","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":98,"snapshot_sha256":"5371d2cb8dc79f5328a2a0dc27bca97865f5a219ae41e4d59d12945441303b88","internal_anchors":4},"formal_canon":{"evidence_count":2,"snapshot_sha256":"3be1c20e0c8aeddc14caf8d6bf0a4756143a8a9703d8217f62d702449ed89def"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}