{"paper":{"title":"GSDrive: Reinforcing Driving Policies by Multi-mode Future Trajectory Probing with 3D Gaussian Splatting Environment","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"A 3D Gaussian Splatting environment probes multiple candidate futures to supply dense rewards that refine end-to-end driving policies.","cross_cats":[],"primary_cat":"cs.RO","authors_text":"Chen Min, Dzmitry Tsetserukou, Shuo Wang, Sifa Zheng, Xuefeng Zhang, Yixiao Zhou, Ziang Guo, Zufeng Zhang","submitted_at":"2026-04-30T16:59:07Z","abstract_excerpt":"End-to-end (E2E) autonomous driving aims to directly map sensory observations to driving actions, but its real-world deployment is hindered by evolving data distributions and the high cost of continual annotation. While combining imitation learning (IL) and reinforcement learning (RL) is a common strategy for policy improvement, conventional RL training relies on delayed, event-based rewards, where policies learn only from catastrophic outcomes such as collisions, leading to premature convergence to suboptimal behaviors. To address these limitations, we propose GSDrive, a framework that uses a"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Evaluated on the reconstructed nuScenes dataset, our method outperforms other simulation-based RL approaches in closed-loop experiments.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The 3D Gaussian Splatting environment provides sufficiently accurate and differentiable simulation of future vehicle dynamics and interactions to produce useful dense shaping rewards that transfer to real-world policy improvement.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"GSDrive combines IL priors with RL feedback by probing multi-mode futures inside a 3D Gaussian Splatting simulator to supply dense rewards for closed-loop driving policy improvement on nuScenes.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A 3D Gaussian Splatting environment probes multiple candidate futures to supply dense rewards that refine end-to-end driving policies.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"733cf257172b3189b6b2a1bf47a1fe2820881f329581bee168d367188accaf94"},"source":{"id":"2604.28111","kind":"arxiv","version":3},"verdict":{"id":"8452bb40-5f93-41a3-ac71-fdaa442894dd","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T16:42:42.310490Z","strongest_claim":"Evaluated on the reconstructed nuScenes dataset, our method outperforms other simulation-based RL approaches in closed-loop experiments.","one_line_summary":"GSDrive combines IL priors with RL feedback by probing multi-mode futures inside a 3D Gaussian Splatting simulator to supply dense rewards for closed-loop driving policy improvement on nuScenes.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The 3D Gaussian Splatting environment provides sufficiently accurate and differentiable simulation of future vehicle dynamics and interactions to produce useful dense shaping rewards that transfer to real-world policy improvement.","pith_extraction_headline":"A 3D Gaussian Splatting environment probes multiple candidate futures to supply dense rewards that refine end-to-end driving policies."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.28111/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_compliance","ran_at":"2026-05-19T18:36:41.646637Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"16646272ba79230c2abdc934ac68516d519d0c6b26c78d656db7c469c766f2d2"},"references":{"count":31,"sample":[{"doi":"","year":2024,"title":"End- to-end autonomous driving: Challenges and frontiers,","work_id":"e86bcc53-963a-47a3-bbfe-dd58108914ba","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"The era of end-to-end autonomy: Transitioning from rule-based driving to large driving models","work_id":"e401a554-5c4f-4b87-8aac-2f54e0f56118","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Iterative label refinement matters more than preference optimization under weak supervision","work_id":"12bd644a-8578-4b18-8fa9-494b3d751a62","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"End-to-end driving with online trajectory evaluation via bev world model,","work_id":"9c7b66de-f25b-4880-b41d-6a56394155d5","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"arXiv preprint arXiv:2503.11650 (2025)","work_id":"d764e87f-6ac5-4436-bca1-d2eca7a1d8a0","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":31,"snapshot_sha256":"f50a50aaab8b82c3819e281ee017a4604326c69858ce4f187bb4a22bd1c18af8","internal_anchors":6},"formal_canon":{"evidence_count":2,"snapshot_sha256":"887639c3c704d66af01e9d9af0ab3c689fff7cf1dd88cee7710542bcdf464dec"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}