History / ICLR 2026

ICLR 2026: create pages for all 39 "(page pending)" papers - Web-verified every pending paper against arXiv/OpenReview and created a page (efficiency, robustness/security, world models, planning, architecture/3D, dexterous/humanoid, navigation, benchmarks, data/memory, tactile, RL) - Replaced all "(page pending)" markers in the ICLR-2026 index and survey with links - 0 papers fabricated; unconfirmable numbers omitted - 0 dangling wikilinks across 263 pages Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Jun 1, 2026

Source-verification audit: fix fabrications, rename PI slugs, link hygiene, add foundational pages - Re-verified all 195 pages vs original sources; removed 47 confirmed fabricated tables/numbers, restored 19 false-positive deletions (full-PDF re-check) - Renamed PI tech-report slugs ICLR-2026-pi07/pi06/RECAP -> PI-pi07/PI-pi06/PI-RECAP (these are PI technical reports, not ICLR 2026 papers); updated 171 wikilinks - Fixed 60 broken wikilinks -> 0 dangling across the wiki - Added foundational pages: OpenVLA, ReKep, AgiBot World Colosseo, RoboBrain 2.0; linked from Home + sidebar + venue indexes Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Jun 1, 2026

Link ICLR-2026 index + survey to the 47 newly-created paper pages Replaced the italic-title placeholders with [[wiki|page]] links for every paper that now has a wiki page. Remaining italics carry the "(page pending)" annotation so a future pass can fill them in. Updated sections in ICLR-2026.md: - Architecture / Spatial / 3D / Efficiency / Robustness / World models / Reasoning / Bimanual & Dexterous / Humanoid / Navigation / RL · Flow / Data & Benchmarks Updated sections in ICLR-2026-VLA-Manipulation-Survey.md: - §1.1 / §1.2 / §1.3 / §2 / §3 / §4 / §5 / §6 / §7 / §8 / §10

Heungwoo committed May 9, 2026

Rebuild ICLR-2026 index from official accepted paper list Source: https://iclr.cc/virtual/2026/papers.html (JSON data file iclr-2026-orals-posters.json). Total: 5,691 accepted papers, ~211 VLA/manipulation related. Changes: - Verified each existing wiki page against the OpenReview decisions via substring matching. 27 pages confirmed accepted; companion technical reports (π0.6, π0.7, RECAP, RLT) and Feb-2026 arXiv preprints (LBM, AsyncVLA, Steerable-Policies) relocated to a separate "preprints / technical reports" section. - Annotated each confirmed page with the official accepted title when it differs from our shorthand (e.g., AutoQVLA → "QVLA: Not All Channels Are Equal", PLD → "Self-Improving VLA via Residual RL", Stage-Aware-RL → "SARM"). - Added ~75 newly-identified accepted papers across architecture, spatial grounding, efficiency, robustness, world models, embodied reasoning, bimanual / dexterous, humanoid / whole-body, navigation, data / benchmarks, and tactile sensing — as candidates for review. - Flagged 9 wiki pages whose paper title could not be found in the ICLR 2026 list (DDVLA, dVLA, DIVA, HiMoE, HyperVLA, Human-Video- Pretraining, OmniSAT, VLA-RFT, XR-1) — likely at NeurIPS 2025 or arXiv-only, pending verification. - Documented the index methodology at the end.

Heungwoo committed May 9, 2026

Steerable Policies: add per-paper page, deepen review, fix 5→6 category error - ICLR-2026-Steerable-Policies.md: new per-paper page (was missing). - Review-Steerable-Policies.md: substantial rewrite of the existing in-depth review: * Fix 5-level -> 6-category error: Tasks / Subtasks / Atomic motions / Points / Gripper traces / Combinations. Prior version under-counted by collapsing Tasks into the trained vocabulary. * Replace generic "Gemini-based labeler" with the actual 4-stage pipeline (Molmo + SAM2 + DETR + Gemini) and report data-scale cascade 38k -> 206k -> ~2M. * Add concrete inference cadences (reasoner N=5, ICL N=20). * Replace bullet-list result summary with detailed Suite A / Suite B tables, baselines table, and explicit note that the two variants are evaluated on different task suites (no head-to-head). * Add Fig. 4 human-oracle finding: unrestricted near-100%, no single category dominates -> structural justification for the 6-category claim. * Architecture: clarify that no architectural surgery is performed, steering commands are plain text tokens in the existing channel. - Home.md / _Sidebar.md / ICLR-2026.md: refresh entries with the corrected vocabulary count, data-pipeline detail, and inference cadence; add per-paper sidebar link. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Heungwoo committed May 6, 2026

Add AsyncVLA (Hirose/Glossop/Shah/Levine, Feb 2026) page + in-depth review - ICLR-2026-AsyncVLA.md: per-paper page on the first hierarchical VLA explicitly designed for seconds-scale cloud-to-edge latency. 8.26B OmniVLA on remote RTX 4090 (5 Hz) + 76M Edge Adapter on Jetson Orin (8 Hz), separated by WiFi with 0.28-6 s delay. 85% success vs. 45%/60% baselines on Vizbot ground-robot navigation. - Review-AsyncVLA.md: 11-section long-form review covering architecture, the two-observation re-conditioning trick, two-stage end-to-end fine-tuning, trajectory re-weighting, full numbers, comparison to RTC / Fast-in-Slow / pi0.7 on the latency axis, and a manipulation port hypothesis. - Home.md, _Sidebar.md, ICLR-2026.md: index/sidebar entries for both pages. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Heungwoo committed May 6, 2026

Add LBM Co-training Study (TRI, Feb 2026) page + in-depth review - ICLR-2026-LBM-Cotraining.md: per-paper page on TRI's controlled 5-modality x 3-strategy ablation (89 policies, 58k sim + 2,835 real rollouts). Verdicts: VL data + cross-embodiment win; discrete action tokens (FAST/VQ-VAE) and CoT-at-inference do not help at LBM scale. - Review-LBM-Cotraining.md: 12-section long-form review covering architecture, all 5 modalities, 3 training strategies, evaluation suite, headline numbers, CoT experiment, discrete-token implications, practitioner recipe, limitations. - Home.md, _Sidebar.md, ICLR-2026.md: index/sidebar entries for both pages. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Heungwoo committed May 6, 2026

Add π0.7 page and π-series evolution comparison π0.7 (Apr 16, 2026) is Physical Intelligence's new steerable generalist VLA, succeeding π0.6. Key novelties: MEM video-history encoder, world- model subgoal-image prompting, episode-metadata + control-mode prompt conditioning with per-component dropout, CFG on metadata, and heavy use of suboptimal / autonomous / RL-rollout data distilled via metadata. Matches RL-specialist performance out-of-the-box and demonstrates the first strong signs of compositional generalization in the π series. New pages: - ICLR-2026-pi07.md — per-paper page - pi-series-evolution.md — model/data/training side-by-side for π0 → π0.7 Updated: - ICLR-2026-pi06.md (marked superseded, forward-link) - ICLR-2026.md, _Sidebar.md, Home.md (navigation) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Heungwoo committed Apr 17, 2026

Add cross-venue RL topic section + two new RL paper pages New top-level topical landing: - RL.md: cross-venue RL-for-VLA topic page with 8 recipe-based subsections (residual RL, RL-token interface, outcome-conditioned, world-model RFT, stage-aware reward shaping, VLM-as-reward, reasoning-token RL, scaling infrastructure) plus adjacent test-time composition and evaluation infrastructure. Includes mermaid recipe-decision diagram, comparison tables, cross-cutting open questions, and a suggested reading order. New paper pages: - ICLR-2026-RL-Tokens: Physical Intelligence's RL Tokens (March 2026 tech report). Tiny RL-token interface bolted onto frozen pi-0.6; tiny actor + critic do online RL in minutes. Up to 3x speedup on screwdriving / zip- tying / ethernet / power-cord; surpasses human teleop speed. Sits with PLD/RFS as the production-engineering branch of residual RL. - ICLR-2026-SimpleVLA-RL: ICLR 2026 paper (arXiv 2509.09674, PRIME-RL). Scalable RL framework for VLA built on veRL with VLA-specific sampling / parallelization / rendering / loss. SOTA LIBERO; beats pi-0 on RoboTwin 1.0 & 2.0; introduces the 'pushcut' phenomenon (RL discovers manipulation patterns absent from SFT data). Restructure: - Home.md: now organized by two axes — Conferences and Topics. RL added as the first topic; tree diagram updated to show paper pages reachable from both axes. - _Sidebar.md: new 'Research topics' section linking to RL; ICLR-2026 RL group cross-references the topic page. - ICLR-2026.md: RL section now points to cross-venue [[RL]] landing. - ICLR-2026-VLA-Manipulation-Survey.md §5: links to [[RL]] topic landing and adds RLT + SimpleVLA-RL rows to the recipe table. - All RL paper pages: back-link footer now reads '← Back to [[ICLR-2026]] · Topic: [[RL]]' so each paper is reachable from both venue and topic axes. Removed: - ICLR-2026-RL-for-VLA.md (consolidated into top-level RL.md) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Heungwoo committed Apr 15, 2026

Add ICLR 2026 VLA & Manipulation survey wiki - Home index + conference landings (ICLR, ICML, CVPR, CoRL, IROS) - ICLR 2026 year landing with categorized paper index - Main survey page: TL;DR, baseline (π0.6), 5 sections, 8 trends, reading list - 33 per-paper pages, each with a mermaid approach diagram + Problem / Method / Results / Significance / Links sections - _Sidebar.md for native GitHub Wiki navigation - _Paper-Template.md for contributors adding future papers Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Heungwoo committed Apr 15, 2026