History

CoRL 2026: expand survey to ~47 confirmed papers + 15 in-depth pages Survey (CoRL-2026-VLA-Manipulation-Survey.md): add all author-announced CoRL 2026 manipulation accepts gathered this session across 8 in-scope themes (WAM/latent-dynamics, VLA, dexterous/tactile, humanoid, RL, policy-adaptation/sim2real, data/human-video, safety) plus adjacent driving/HRI/locomotion; new §2.10 index of in-depth pages; refreshed §1/§3 with the "prediction-as-execution-monitor" and "reactive-inside-the-chunk" insights. All entries source-verified against arXiv; acceptance marked ✅ author-confirmed / ⚠️ reported. 15 new in-depth per-paper pages (Problem/Method/Results/Why/Limitations, arXiv links + one paper figure each, all facts verified against source): SG-WAM, K-UBM, StressDream, MolmoAct2, MolmoBOT, VLA-Feedback, SAE-VLA, FTP-1, HiPHI, CHIP, TOPReward, VLS, HuRo, LUCID, World-In-Your-Hands. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Sep 28, 2026

Add CoRL 2026 survey (preliminary, community-sourced; official list pending) CoRL 2026 (Austin, Nov 9-12; 687 accepted / 32.8% / 2,094 submitted; official per-paper program not yet public). Preliminary manipulation-centric survey with explicit acceptance-confidence marking: - Confirmed CoRL 2026: SG-WAM (geometry-aware WAM), FiberTune (robustness- preserving VLA fine-tune), Dex-X (visual-tactile from human video via sim), Touch2Trace (tactile IL, cable tracing). - Reported/unverified: StellaVLA (in-context VLA), Choice Policies (Berkeley/ Malik, whole-body humanoid), Weave (whole-body dexterous loco-manip). - Excluded ManiFlow (it's CoRL 2025, not 2026). Prominent caveat + confidence legend; to be promoted to a full session-taxonomy survey when the program publishes. Linked from CoRL hub, Home venue table, sidebar. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Sep 28, 2026

Split Reviews into topic-reviews (Reviews) + per-paper long-forms (Reviews-Per-Paper) Durable fix for the recurring 'too long to render' on the growing catalog: - Reviews.md keeps cross-paper topic reviews, lab/series programs, latest-paper reviews (78 links). - New Reviews-Per-Paper.md holds all single-paper long-forms — architecture/ runtime, data/training, world-models/tactile, hybrid-MoT, dexterous-hand data, multi-task/in-context, IROS 2026 full-paper analyses, RSS pointer (56 links). - Both now safely under GitHub-wiki's ~100-link render limit. - Linked Per-Paper from Home and the sidebar; added the IROS 2026 + hybrid + dexterous-data long-forms that weren't catalogued before. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Sep 12, 2026

Fix Reviews (In-depth reviews) 'too long to render': cut duplicate RSS per-paper link block Reviews.md crept to 103 wikilinks (over GitHub-wiki's ~100 catalog render limit) as IROS/in-context sub-entries were added. The 'RSS 2026 per-paper pages' block (16 links) duplicated the RSS survey's own catalog, so replaced it with a one- line pointer to the RSS survey + RSS-2026-Papers index. Total links 103 -> ~88, safely under the limit. No unique navigation lost (RSS pages reachable via the survey). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Sep 7, 2026

IROS 2026: add arXiv links + paper figures to PT / FILIC / Articulated-Tools / DreamMimic Found arXiv preprints and embedded each paper's architecture/pipeline figure (with attribution): Proprioceptive Transformer (2605.21330), FILIC (2509.17053), Articulated-Tools Sim2Real (2509.23075), DreamMimic (2608.22278). MrGrasp and DexKP-VLA have no locatable arXiv preprint (pre-conference) — left with program links only. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Sep 7, 2026

IROS 2026: add in-depth pages for the 5 remaining technical themes (full-coverage complete) Extracted representative papers for the previously session-level-only themes and wrote abstract-verified in-depth pages: - MrGrasp (#2149, HUST) — multi-rate VTLA for fragile deformables (covers deformable + tactile-in-loop) - Proprioceptive Transformer (#3215, ETH Zurich, award finalist) — joint-sensing- only in-hand cube rotation, 3.1x speed - Articulated-Tools Sim2Real (#3176, UCSD/Yip) — sim base + hardware-demo tactile refinement for articulated tools (sim-to-real robustness) - FILIC (#1885, Tsinghua) — dual-loop force-guided IL + impedance, force-aware without F/T sensor (force-from-demonstration) Survey §3.6 representative table now spans 17 technical themes; coverage note updated to reflect full technical-space coverage. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Sep 7, 2026

IROS 2026: full-coverage representative selection + 4 new in-depth pages + WAM-vs-VLA trend update - Added §3.6 'Representative papers by technical theme' to the survey with an explicit selection criterion (technical/methodological distinctiveness, NOT lab/vendor prestige) — a 12-theme table spanning efficiency, RL, WM-as-scaffold (3 modes), bimanual (2), in-context (2), memory, dexterous grasping (2). - New abstract-verified in-depth pages (official program): * DreamMimic (#279, SJTU/Tsinghua) — WM as humanoid distillation scaffold * Scaling Cross-Embodiment WM (#2451, SJTU/UCSD-Hao-Su/MIT-Yilun-Du) — particle WM as shared cross-embodiment interface + MPC * DexKP-VLA (#752) — VLM->keypoints->optimizer, zero-shot dexterous grasping * VCoT-Grasp (#1923) — end-to-end grasp FM + visual chain-of-thought - WAM-vs-VLA TREND UPDATE (genuinely new): across IROS 2026's WAM papers the WM is pushed out of the inference loop into a training/representation SCAFFOLD (critic / distillation teacher / shared-representation+MPC) around a reactive VLA — the 'versus' dissolves. Added to survey §5.1 and Review-WAM-vs-VLA- Robustness §11. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Sep 7, 2026

Fix 'too long to render' on Review-World-Models and Reviews: split link-dense mega-lines Both pages hit GitHub-wiki's render budget after recent growth. Root cause was single lines packed with 15-17 wikilinks (O(n^2) emphasis/link parsing) — the same trigger as the earlier RSS survey. Split the pure link-list lines into shorter lines (content and all links preserved; max wikilinks-per-line 17->7 in World-Models, 15->6 in Reviews). No links removed. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Sep 7, 2026

Add IROS 2026 to navigation: IROS venue hub (IROS.md) + sidebar The IROS 2026 survey existed but the venue navigation (IROS.md Surveys section and _Sidebar.md) only listed IROS 2025. Added IROS 2026 survey link (with its full-paper analyses) to the IROS hub, and an IROS 2026 sidebar entry. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Sep 7, 2026

IROS 2026 survey: add navigation links (out-of-scope note in §2 + nav/mobility links in §7) Navigation was absent from the manipulation-scoped survey. Added a §2 out-of- scope note flagging IROS 2026's navigation cluster (visual nav, social nav, drone nav, world-model-for-nav) and the manip<->nav crossover, plus a §7 Navigation & mobility link row (Qwen-RobotNav, Embodied-Nav-FM, OmniVLA-Nav, CompassNav, CE-Nav, Executable-3DGS-Nav; crossover: Joint Nav+Manip Planning, Humanoid loco-manipulation). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Sep 7, 2026

In-context imitation: add in-depth pages for ICRT, Behavior Prompting, MimicDroid (w/ paper figures + arXiv) - Review-ICRT (arXiv 2408.15980, ICRA'25, Berkeley): causal-transformer next- token ICL over sensorimotor tokens — the cluster-E anchor; method figure. Framed as the reference the newer papers improve on (SSM/visual-reasoning/data). - Review-Behavior-Prompting (arXiv 2606.30457, Stanford/Shuran Song): single-demo 'behavior prompt' via cross-attn prompt encoder + diffusion decoder; key finding = task diversity drives prompting; iPhUMI handheld data + DrawAnything/LIBERO-Gen. Architecture figure. - Review-MimicDroid (arXiv 2509.09769, ICRA'26, UT Austin RPL): learns ICL from unlabeled human play video (no teleop) via similar-behavior pairs + wrist-pose retargeting + patch masking; ~2x real success. Overview figure. - Linked all three from Review-In-Context-Imitation (clusters E/F + comparison table). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Sep 7, 2026

IROS 2026 major papers: upgrade to in-depth with paper figures + arXiv links; add VLA-RL page - Added arXiv links + embedded the papers' own architecture/results figures (with attribution) to: AtomVLA (2603.08519, 2-stage WM-critic pipeline), 3D FlowMatch Actor (2508.11002, 3D-scene-token architecture; +21d->16h / 0.5->18.2Hz), RoboSSM (2509.19658, results: SSM holds vs ICRT collapse as demos->32, 16x longer prompts), TempoFit (2603.07647, layer-wise KV-memory method figure). - New IROS-2026-VLA-RL page (2505.18719, Tsinghua/NTU): trajectory-as- conversation online RL + VLM process-reward; OpenVLA-7B +4.5% LIBERO, real 60->90%, inference-scaling; framework figure embedded. Linked from survey and Multi-Task-VLA review. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Sep 7, 2026

VLA Hybrid review: reflect IROS 2026 — add AtomVLA critic-only mode (3rd Axis-2 answer) Insight reinforced not overturned: convergence continues, latent/low-inference- cost corner keeps winning. Refinements: - Axis 2 gains a THIRD option: 'training-time critic only' (AtomVLA) — WM neither in-path nor a co-training tower; it scores action chunks during offline GRPO then is absent at deploy. - Added AtomVLA to structural-variants table + §4c latency-frontier bullet. - New §4f 'IROS 2026 — what changed': critic-only mode + broadened WAM roles (cross-embodiment WM, DreamMimic, RoboDream data-factory); marginal-value gap still stands. - Refined open-question 3 (in-path vs reactive vs critic-only, task-dependent). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Sep 5, 2026

IROS 2026: add full-paper pages for RoboSSM and TempoFit; link from survey + reviews - IROS-2026-RoboSSM (KAIST + UT Austin/Peter Stone): scalable in-context imitation via state-space models (Longhorn SSM, linear-time, length extrapolation, LIBERO, open-source). The in-context/memory efficiency answer. - IROS-2026-TempoFit (XJTLU): training-free temporal retrofit that reuses a frozen VLA's prefix-attention K/V as content-addressable memory (layer-wise FIFO K/V + Frame-Gap Temporal Bias); LIBERO-Long +4.0%, near-real-time. - Linked both from the IROS survey (§3.3, §5.2), Review-In-Context-Imitation, and Review-VLA-Memory (structured-vs-parametric row). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Sep 5, 2026

IROS 2026: add 5 full-paper analyses + reflect into existing in-depth reviews Full-paper pages (abstract-verified from official program): - IROS-2026-AtomVLA: subtask-aware VLA + latent-WM scoring for offline GRPO (WAM lens) - IROS-2026-3D-FlowMatch-Actor: CMU/NVIDIA unified single/dual-arm 3D policy, +41.4% PerAct2, ~30x faster (bimanual SOTA) - IROS-2026-EquiBim: symmetry-equivariant bimanual policy - IROS-2026-IMLE-VLA: single-step cIMLE action head, 55Hz, LIBERO 98.0% (efficiency) - IROS-2026-ICLR-Visual-Reasoning: in-context imitation with image-space reasoning traces Reflected IROS 2026 into existing reviews: - Review-World-Models: WAM-as-critic row (AtomVLA offline GRPO) - Review-In-Context-Imitation: ICLR-visual-reasoning + RoboSSM - Review-VLA-Memory: structured-vs-parametric memory row (GaussMemory/PROMPT vs TempoFit/RoboSSM) - Review-Multitask-VLA: VLA-RL / LAR-MoE / AtomVLA / MoE-humanoid - Review-Realtime-Execution: single-step head row (IMLE-VLA) - Review-Humanoid-VLA: IROS bimanual/whole-body trend (3DFA/EquiBim/ULTRA/CEER/MoE-VLA) Linked all from the IROS 2026 survey. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Sep 5, 2026

IROS 2026 survey: add WAM / memory-in-context / bimanual-humanoid trend deep-dives; abstracts are program-available - Corrected sourcing note: the program's local search exposes titles+authors+ abstracts (abstract-informed, program-verified) — not titles-only. - New §5 trend deep-dives (three requested lenses): * WAM: AtomVLA (latent WAM post-training), Scaling Cross-Embodiment World Models for Dexterous Manip, DreamMimic/Touch-Dreaming (humanoid), RoDyn, RoboDream (data factory), GrndCtrl — insight: latent/deployable WAM ascendant. * Memory & in-context: dedicated 'Building Memory into VL Robot Agents' session; GaussMemory/PROMPT/TempoFit (structured vs KV memory); RoboSSM (SSM in-context), ICLR (visual-reasoning in-context); IMLE-VLA/ESPADA demo-efficiency — insight: structured vs parametric memory bifurcation. * Bimanual/humanoid: EquiBim, 3D FlowMatch Actor, HapticVLA, Grasp-Handover- Rotate, ULTRA/CEER/OmniDP whole-body, MoE-VLA humanoid, SPIDER retargeting — insight: coordination structure + data bottleneck. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Sep 5, 2026

Add IROS 2026 VLA & Manipulation survey (official program, pre-conference) IROS-2026-VLA-Manipulation-Survey: built from the public official program (2026.ieee-iros.org, 1,900+ papers, Pittsburgh Sep 27-Oct 1). Session-level taxonomy of ~40 in-scope manipulation/VLA/dexterous/imitation sessions + sampled papers extracted per theme (VLA: AnyCamVLA/LangGap/OG-VLA/BFA++/ Safe-Night-VLA; scaling/RL: VLA-RL/LAR-MoE/flow-matching; in-context imitation: RoboSSM/ICLR-visual-reasoning/IMLE-VLA/ESPADA; policy: MaskVLA/SynthLA; language-in-loop: NL2SpaTiaL/AURORA). Clearly flagged title-level (abstracts pending Xplore post-conference), not abstract-verified. Added to Home venue table + top pointer. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Sep 5, 2026

Add in-depth survey: In-Context Imitation & Demo-Following Review-In-Context-Imitation: watch-a-demo-and-reproduce-it (no per-task FT) as a memory-conditioning problem. Taxonomy by how the demo is conditioned — (A) cross-attention/video-conditioned (Vid2Robot, VLBiMan, See-Once-Then-Act), (B) recurrent/query memory (HAMLET, MemoryVLA, RememVLA, ContextVLA), (C) fast-weight/TTT (RoboTTT), (D) retrieval (MemER, MAP-VLA, Memory-Retrieval, KEMO, Long-Context-IL), (E) token-sequence ICL (ICRT, Behavior Prompting), (F) play-video ICL (MimicDroid). Mapped to RoboMME's Imitation (procedural memory) suite; comparison table + design axes + open challenges. Cross-linked from Reviews and Home. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Aug 31, 2026

Add in-depth survey: Egocentric Video for VLA Pre-Training Review-Egocentric-Video-Pretraining: how label-free first-person human video becomes a pretraining signal. Taxonomy of methods (A pseudo-action extraction: hand-pose/keypoints/optical-flow/part-motion; B latent-action models; C world-model/video-prediction; D reconstruct-then-retarget; E auxiliary-modality recovery), the dataset landscape (EgoDex, EgoVerse, EgoScale, Being-H0, DYNA-2, DreamDojo, UniDex, EgoVLA), scaling-law evidence (EgoScale R2=0.9983 +54%, DYNA-2 transfer law), the transfer question (emergence/decoupling/synthesis + two-stage recipe), and open challenges. Cross-linked from Reviews and Home. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Aug 25, 2026

Add in-depth review of NVIDIA RoboTTT (context scaling via TTT in GR00T N1.7) Review-RoboTTT (arXiv 2607.15275, NVIDIA GEAR + Stanford + UT Austin): 8K-timestep visuomotor context at constant latency by adding Test-Time- Training fast-weight layers to GR00T N1.7's DiT action head. Detailed GR00T implementation section: TTT layer after self/cross-attn in each of 16 DiT layers (~10M each -> ~690M), tanh-gated to preserve pretrained skills; register tokens (N=16) carry compressed VL history through TTT while VL tokens bypass; fast weights = 2-layer GeLU MLP updated per step (W_t <- W_{t-1} - eta*grad MSE(f(K),V)), read via Q; training recipe = flow matching + sequence action forcing (per-step tau) + TBPTT (fast weights carried, gradients detached at segment boundaries); 30 Hz on RTX 5090 (YAM bimanual). Results, new capabilities (one-shot in-context video imitation, DAgger- distillation self-improvement, perturbation robustness), limitations. Cross-linked from Reviews, Home lab programs, and Review-GR00T-Series. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Aug 19, 2026

Add in-depth Review pages for HiMoE-VLA and DyGRO-VLA (multi-task VLA cluster) - Review-HiMoE-VLA: hierarchical depth-wise MoE (shallow AS-MoE per action space, deeper HB-MoE per embodiment, central dense consolidation; AS-Reg/ HB-Reg; 32 experts top-4). Negative-transfer ablation (dense pi0 -0.259 vs HiMoE +0.186), LIBERO 97.8 / real xArm7 75.0 / ALOHA 63.7. Framed as the MoE-routing cluster of multi-task fixes. - Review-DyGRO-VLA: cross-task RL scaling — protect a shared latent (offline info-theoretic pretrain) then optimize grouped RL residuals (a=Δa+a_base, K-critic ensemble, entropy load-balance). LIBERO 92.7->97.1, Long 85.2->95.0. - Repointed Multi-Task VLA links to the new Review pages; added in-depth pointers to the ICLR/ICML venue entries; catalog entries in Reviews. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Aug 14, 2026

Add Multi-Task VLA review + MergeVLA page - Review-Multitask-VLA: why one VLA fails across many tasks (6 failure modes: negative transfer/gradient conflict, non-mergeability, multi-task conflict, catastrophic forgetting, routing confusion, instruction collapse) + a 7-cluster solution landscape (merging, MoE, gradient control, skill decomposition, instruction grounding, continual, adapters) with a decision guide and open questions. Anchored on MergeVLA (CVPR 2026). - CVPR-2026-MergeVLA: per-paper page (non-mergeability diagnosis + task-masked LoRA / cross-attention-only action expert / test-time task router). - Cross-linked from Reviews catalog and Home topic reviews. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Aug 14, 2026

Pyramid: split into two tables (papers vs companies); mark gripper-based within each - §3a Research papers and §3b Companies are now separate tables (same columns). - Each table has a 🖐 five-/multi-finger group and a 🤏 gripper/2-finger group. - Gripper-based companies kept IN the company table but clearly marked 🤏: Physical Intelligence (parallel-jaw gripper), Dyna/DYNA-2 (gripper-primary; claims 5-finger, unverified), AgiBot/Zhiyuan (modular gripper). YUBI stays as the 🤏 gripper paper. Removed the standalone §3d gripper-reference (folded into the two tables); industry-split renumbered to §3d. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Aug 14, 2026

Pyramid: merge papers+companies into one grouped table; split out gripper/non-5-finger setups - Combined §3 into a single table with '📄 Research papers' and '🏢 Companies' divider rows (papers and companies grouped but distinguished, same columns); narrative moved below (§3b verdict, §3c industry split). - Per user: removed non-5-finger setups from the dex-hand table — Physical Intelligence (π, parallel-jaw gripper generalist) and YUBI (2-finger bidigital handheld) — into a new §3d 'gripper / non-5-finger (reference)' table for separate comparison, with a data-speed-vs-dexterity takeaway. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Aug 14, 2026

Split pyramid table into papers (§3) vs companies (§3c); fix Genesis sim=eval-only - §3 is now research papers/methods only; §3c is companies/industry programs (Genesis, DYNA-2, RLDX-1, GR-Dexter, PI, Figure, Tesla, AgiBot, Unitree, Fourier, teleop-tool vendors) — both use identical L1-L6/tactile/teleop-FT columns so they read consistently. - FACT-CHECK FIX (per user): Genesis GENE-26.5 uses NO simulation training data — pretrain is real human data only (glove + ego + third-person video, >200k h); Genesis-World sim is closed-loop EVAL only ('zero simulation training data'). Corrected L5 from ● to △ (eval-only) in §3c, fixed §3b, and corrected the Review-Genesis-GENE page (TL;DR, data-engine table, pyramid placement, significance, links). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Aug 14, 2026

Dex-Hand Data Pyramid: add 3c industry data strategies (Figure, Tesla, AgiBot, Unitree, Fourier) New 3c company-data-strategy table + analysis: hardware incumbents are teleop/fleet-first (Figure 10Gbps fleet offload + 3g fingertip tactile; Tesla Optimus Gen-3 22-DoF, factory data; AgiBot World >1M free-form teleop trajectories / 4000m2 facility / GO-1; Unitree open full-body; Fourier ActionNet ~140h) vs newcomers (Genesis/Dyna/RLWRLD) shrinking L6 to a sub-hour tip. Notes the China-vs-US open-vs-proprietary split. Updated 5 and 7 accordingly. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Aug 14, 2026

Add in-depth page for Genesis AI GENE-26.5 (glove-first dexterous FM, vendor) Review-Genesis-GENE: Genesis AI's May 2026 launch. 1:1:1 tactile-glove <-> human-hand <-> robot-hand co-design (collapses L4 retargeting), data engine (glove demos + ego + internet video + simulation), sub-hour robot fine-tune (secondary-sourced). Flagged throughout as vendor claims: no paper/weights/ benchmarks, undisclosed hand DoF and architecture, unquantified capabilities. Linked from the Dex-Hand Data Pyramid (matrix, 3b, links) and Reviews catalog. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Aug 13, 2026

Dex-Hand Data Pyramid: re-center matrix on training-necessity + add recent (Genesis, GR-Dexter, Dexterous Point Policy) - New legend separates training-necessity from eval-only: ● required / ○ auxiliary / △ eval-or-deploy-only (NOT training) / blank. - Added 'Teleop -> fine-tune?' column answering directly whether target-hand teleop/robot data is used for training (Yes+volume / eval-only / No). - Re-audited every row: foundation models fine-tune on a small teleop tip (● L6), while device-free-video + capture-interface methods use the robot only for eval/deploy (△). Fixed prior conflation of eval with training. - New rows: Genesis GENE-26.5 (glove-first, <1h robot FT, vendor), GR-Dexter (ByteDance, ~20h/task trains VLM+action DiT), Dexterous Point Policy (KAIST, zero-teleop-training, robot=eval-only 75% vs 1%). - New 3b analysis + verdict; updated 5/6/7 (teleop = shrinking <=1h tip). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Aug 13, 2026

Fix RSS survey 'too long to render' (recurrence): split Themed reading list to its own page Prior bullet-split left the survey at 147 wikilinks + six ~950-char prose paragraphs — still near GitHub-wiki's render budget, so the error recurred. Move the link-dense Themed reading list (94 of 147 links) to a dedicated RSS-2026-Reading-List page (same pattern as the earlier session-tables -> RSS-2026-Papers split), leaving a pointer. Main survey now 54 links; new page 101 links, both with short per-line link density. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Aug 13, 2026

Add in-depth pages for T-Rex and RLDX-1 (tactile-reactive + industry dexterity FM) - Review-T-Rex (arXiv 2606.17055): variable-rate Mixture-of-Transformers, slow Action Expert + fast Tactile Expert; temporal tactile VQ-VAE; Dexmate Vega-1 + 2x Sharpa Wave 22-DoF, 5 fingertip sensors, 300 Hz; 12 tasks 65% avg vs EgoScale 35% (+30pts); 100h/22-primitive dataset. - Review-RLDX-1 (RLWRLD, Seoul, vendor): dexterity-first foundation model, Multi-Stream Action Transformer (vision/motion/memory/torque), bare- human-hand + five-finger retargeting (>200 demos/hr) + synthetic aug; ALLEX/Franka/OpenArm; vendor benchmarks vs pi0.5/GR00T N1.6 (flagged as company claims, not peer-reviewed). - Linked both from the Dex-Hand Data Pyramid (matrix + tactile/industry deep-dives) and the Reviews catalog. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Heungwoo committed Aug 13, 2026