Ad Literature Survey

World Models

• NeurIPS / ICLR / ICML — dominant venues (world model architectures, video prediction)
• CVPR — vision-centric world models, video generation for AD
• CoRL — world models applied to robot/driving policy learning

VLA / VLM Models

• CVPR / ECCV / ICCV — vision-language grounding, embodied VLMs
• NeurIPS / ICLR — foundation model architectures, large-scale training
• CoRL — VLA for robotic action, end-to-end control
• CVPR WAD workshop — AD-specific VLM/VLA applications

RL Methods (AD-focused only)

• CoRL — RL-based driving policies, sim-to-real transfer
• RSS — safety-constrained RL, constrained MDPs for AD
• IEEE IV / ITSC — RL for traffic scenarios, intersection handling, merging
• CVPR WAD / NeurIPS ML4AD workshops — RL applied to end-to-end driving

Prediction & Planning

• CVPR / ICCV — motion forecasting benchmarks (Argoverse, nuScenes)
• NeurIPS — trajectory prediction, probabilistic forecasting
• ICRA / IROS — planning algorithms, motion primitives
• CoRL — learned planners, imitation + RL
• RSS — formal planning with safety certificates

Key Journals

Key Workshops

Key Benchmarks & Datasets (for grounding paper claims)

• nuScenes — prediction, planning, occupancy, VLM perception
• Waymo Open Dataset — motion forecasting, 3D detection
• Argoverse 1/2 — trajectory prediction, map-aware planning
• CARLA — RL training, world model evaluation, sim-to-real
• nuPlan — reactive closed-loop planning evaluation
• DriveLM / DriveVLM — VLA/VLM evaluation for AD

Search Strategy

Phase 1: Scope Definition

1. Identify the core research question and AD subtopic
2. Map to relevant venues from the lists above
3. Identify time range (field evolves rapidly — default to last 3 years, extend if tracing lineage)

Phase 2: Systematic Search

1. Primary databases: IEEE Xplore, Scopus, Google Scholar, Semantic Scholar, arXiv (cs.RO, cs.CV, cs.AI, cs.LG, cs.MA)
2. AD-specific repositories: nuScenes papers, Waymo Open Dataset papers, KITTI benchmark papers
3. Standards search: ISO catalogue, SAE MOBILUS, UNECE document database

Phase 3: Keyword Expansion

AD research uses evolving terminology. Always expand searches with:

Phase 4: Synthesis

1. Organize findings by methodology, chronology, or taxonomy
2. Identify research gaps and contradictions
3. Map the citation network — find seminal papers and recent extensions
4. Note which research groups lead each subtopic
5. Highlight methodological trends (sim-to-real, data-driven, formal methods, etc.)

Key Research Groups (Non-exhaustive)

World Models

• NVIDIA Research — GAIA-1 lineage, neural world simulators, DriveX
• Waymo Research — large-scale world modeling, closed-loop simulation
• Waabi — world models for scalable AD training (Prof. Urtasun)
• Toyota Research Institute — generative world models for safety-critical evaluation

VLA / VLM Models

• Shanghai AI Lab / OpenDriveLab — DriveLM, DriveVLM, UniAD, SparseDrive
• Tsinghua University (MARS Lab) — VLM-based driving, language-conditioned planning
• UC San Diego / UCSD — LLM/VLM grounding for embodied agents
• Microsoft Research — DriveGPT, multimodal driving foundation models

RL Methods (AD)

• UC Berkeley / BAIR — RL for driving, offline RL, decision transformers
• Stanford SAIL — safe RL, constrained policy optimization
• ETH Zurich (Prof. Frazzoli / Zeilinger) — safe RL, MPC-RL hybrid methods
• Wayve — end-to-end RL for real-world AD
• NTU / Chen Lv group — RL for AD decision-making, human-machine interaction (key authors: Zhiyu Huang, Jingda Wu, Haochen Liu, Xiaoyu Mo)

Prediction & Planning

• CMU / Argo AI Lab — motion forecasting, Argoverse benchmark
• ETH Zurich (Prof. Geiger / Oswald) — trajectory prediction, social dynamics
• MIT CSAIL — scene-aware prediction, occupancy forecasting
• nuTonomy / Motional — nuScenes benchmark, map-aware planning
• UCLA Mobility Lab / Jiaqi Ma group — embodied AI planning, perception-prediction-decision integration (key authors: Zhiyu Huang, Yuxiao Chen, Zewei Zhou)
• NVIDIA Research AV Group — generative planning, interaction-aware prediction (key authors: Boris Ivanovic, Marco Pavone)
• UC Berkeley / Masayoshi Tomizuka group — learning-based planning, human-robot interaction, sim-to-real (key authors: Zhiyu Huang, Bolei Zhou)

Output Format

For literature review results, provide:

1. Summary table — one row per paper: | Author | Year | Venue | Method | Benchmark / Metric | Key Finding | Limitation | Code | Citation |

2. Quantitative comparison table — side-by-side numbers on shared benchmarks (e.g., minADE/minFDE on nuScenes, PDMS on nuPlan, WOMD soft mAP). Only include benchmarks where ≥3 papers report results.

3. Taxonomy/categorization — group papers by approach (e.g., regression-based vs. generative, model-based vs. model-free RL, autoregressive vs. diffusion world models).

4. Gap analysis — identify under-explored areas and contradictions across papers.

5. Draft related work paragraph — 3–5 sentences grouping papers by approach, ready to paste into a paper. Example structure: "Early works on X used ... [cite]. More recent approaches adopt ... [cite]. However, these methods share the limitation of ..."

6. Recommended reading list — 3–5 essential papers + 5–10 extended reading, with one-line justification for each.

Note: Timeline is optional — only include if explicitly asked or for survey-style requests.

Citation Rules

• NEVER fabricate content — all claims in paper reviews (limitations, methods, results, datasets) must be extracted directly from the paper. Do not infer, generalize, or add context not present in the source.
• NEVER hallucinate paper titles, authors, or venues
• When searching programmatically, verify via Semantic Scholar API or Google Scholar
• For standards, always cite the specific edition/year (e.g., ISO 26262:2018, not just "ISO 26262")
• Use the citation format required by the target venue (typically IEEE for AD conferences)