Key Papers and Blogs — The Canonical Reading List

The ~50 artifacts (papers, blog posts, talks) that this repo's docs keep citing, consolidated in one place with a one-line "why it matters" each. Organized as: start-here essentials, then by theme, then the staying-current feeds. Time estimates assume focused skimming, not line-by-line study.

Conventions: ★ = read in full, the rest are skim-the-right-sections. Each entry notes which doc in this repo gives it context.

Last verified: 2026-06-12.

---

0. If you read only ten things

That set covers the four loops, both verticals, and the central controversy (do benchmarks measure anything?). Everything below is depth.

---

1. The data engine, labeling, and curation

(context: 04-labeling-and-data-curation.md)

• Auto4D (arXiv:2101.06586) — the canonical academic auto-labeling reference: 4D boxes from point-cloud sequences.
• Offboard 3D Object Detection (arXiv:2103.05073) — Waymo's non-causal teacher detector; why offline labels beat online perception.
• Noisy Student (arXiv:1911.04252) — still the cleanest self-training reference.
• LIMA (arXiv:2305.11206) + DataComp-LM (arXiv:2406.11794) — the LLM-side evidence that curation beats scale; the same lesson AV/robotics converged on.
• Constitutional AI (arXiv:2212.08073) — the conceptual progenitor of RLAIF; the bridge from LLM alignment to preference labels for driving/manipulation.
• Grounding DINO + Grounded-SAM 2 — repos, not papers; run them (project 03).
• Mobileye REM technology page — the cleanest large-scale crowdsourced data flywheel; study the semantic-compression trick.
• Meta EgoBlur — the open PII-redaction standard (project 06).
• Blogs to skim quarterly: Voxel51, Encord — current tooling vocabulary lives here.

2. World models and generative video

(context: 05-world-models-and-generative.md)

• Ha & Schmidhuber, "World Models" (arXiv:1803.10122) + DreamerV3 (arXiv:2301.04104) — the RL grounding, so the term can't bluff you.
• GAIA-1 (arXiv:2309.17080) — the field-defining driving WM.
• Cosmos-Transfer1 (arXiv:2503.14492) — the controlnet-style head behind the hybrid-sim pattern (project 09).
• Cosmos-Drive-Dreams (arXiv:2506.09042) — "generate from labels, don't label generations"; comes with an open dataset.
• V-JEPA 2 (arXiv:2506.09985) — the LeCun camp's strongest result: latent prediction → zero-shot robot planning from 62 robot-hours.
• Genie 3 blog (DeepMind) — the interactive-WM frontier: real-time, multi-minute, promptable worlds.
• VideoPhy (arXiv:2406.03520) + PhyGenBench (arXiv:2410.05363) + WorldModelBench (arXiv:2502.20694) — the reality-check benchmarks for "does it understand physics."
• Waymo World Model blog (Feb 2026) — Waymo's Genie-3-based entry; the strongest "WM as simulator" claim from a fleet operator.
• The debate itself: LeCun's JEPA argument (talk) vs the Cosmos paper's scale position — hold both.

3. VLAs and robot learning

(context: 02-robotics-foundation-models.md)

• RT-2 — actions-as-text-tokens; where VLA semantic generalization came from.
• OpenVLA (arXiv:2406.09246) — the open 7B baseline you can actually fine-tune (project 11).
• π0-FAST research page — frequency-space action tokens; why autoregressive VLAs got 5× cheaper to train.
• π0.7 blog (Physical Intelligence) — compositional generalization; the strongest late-2025 open-recipe signal.
• Gemini Robotics 1.5 / ER 1.5 blog — the hierarchical thinking-VLM + acting-VLA pattern.
• Figure Helix + Helix 02 — System 1/2 (and 0) humanoid architecture, explained by the people shipping it.
• GR00T N1 (arXiv:2503.14734) + GR00T-Dreams — the synthetic-data-trained humanoid foundation model; the "36 hours vs 3 months" claim, with caveats.
• DROID (arXiv:2403.12945) — what deliberate diversity curation looks like in robot data collection.
• MimicGen — the 10-demos → 1000s-of-demos replay multiplier; the economics fix for teleop cost.
• ALOHA / Mobile ALOHA — the low-cost teleop rig that reset data-collection economics.
• LeRobotDataset v3.0 blog — the de-facto open robot-data format; read before designing any robot-data pipeline.

4. AV perception, prediction, and closed-loop eval

(context: 01 + 06-open-problems-and-benchmarks.md)

• BEVFormer (arXiv:2203.17270) — the canonical multi-camera BEV architecture (project 04).
• CODA (arXiv:2203.07724) — the corner-case benchmark behind the "<12.8% mAR" headline; perception is unsolved on the long tail.
• MultiPath++ (arXiv:2111.14973) — the classic motion-forecasting baseline (project 13).
• EMMA (Waymo, arXiv:2410.23262) — Gemini-based end-to-end driving; the openly-described counterpart to Tesla v14.
• NAVSIM v2 + nuPlan-R (arXiv:2511.10403) — the current closed-loop eval arenas; nuPlan-R adds diffusion-based reactive agents.
• WOSAC 2025 winner SMART-R1 (arXiv:2509.23993) — where sim-agents realism actually stands (≈0.786, barely moving).
• Waymo Open Dataset E2E benchmark (arXiv:2510.26125) — curated long-tail E2E driving with human Rater Feedback Score; the methodological move away from log-replay.
• Wayve PRISM-1 + EmerNeRF + SplatAD (arXiv:2411.16816) — the real-to-sim reconstruction lineage (project 08).

5. Strategy, industry, and safety framing

(context: 03, 06, 09)

• Applied Intuition 2025 Year in Review — the company's own statement of product surface and scale.
• Wayve × Sequoia "AV 2.0" interview — the E2E worldview from its most articulate proponent.
• TIME: How the Meta–Scale deal upended the data industry — why "neutrality" became a data-vendor differentiator overnight.
• TechCrunch: the rise and fall of Tesla Dojo — the cleanest evidence that compute is not the bottleneck, data is.
• ISO 21448 (SOTIF) — read a good secondary explainer rather than the standard itself first; then skim the standard's clause structure. Pair with the UN R157 amendment notes.
• ASAM OpenSCENARIO DSL spec — first 20 pages — scenario-language literacy; customers grade vendors on this.
• Pebblous post-mortem: the rise and fall of synthetic-data companies — why pure-rendering startups died in 2025.

---

6. Staying current — the feeds that matter

(The full operating manual is 09 §F; this is the condensed version.)

Daily (≤30 min): Hugging Face Daily Papers — the single highest-signal feed in 2026.

Weekly:

• Wayve Thinking — the best industry blog on driving WMs and embodied AI.
• NVIDIA Developer blog — Robotics/Cosmos/DRIVE tags — release notes that matter.
• Import AI — cross-cutting weekly.
• The Robot Report — funding/deployment signal.
• Physical Intelligence blog — sparse; every post matters.

Quarterly: re-skim OpenDriveLab (challenges/benchmarks), Applied Intuition newsroom, Voxel51 + Encord blogs, and the accepted-paper lists of the nearest of CVPR / CoRL / NeurIPS / ICRA using the filter-triage-deep-read flow in 09 §F.4.

---

7. How to use this list

1. Week 1: the "only ten things" table, in order. That is ~7 hours and covers 80% of conversations.
2. Weeks 2–4: pull the theme section matching whichever project (01–20) you're working through — the projects' own "further reading" sections go deeper per topic.
3. Ongoing: the §6 feeds. When something big lands (a Cosmos/GAIA/π release, a new eval result), trace it back to the theme section here and ask which entry it supersedes.
4. Every 6 months: prune. An entry that no longer earns its place (superseded model, dead benchmark) gets replaced — this list is a living artifact like the rest of the repo (see 08 §G).