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>
Add in-depth review of DuoCore-FS (Astribot async whole-body VLA) Review-DuoCore-FS.md (624 lines). Paper: Asynchronous Fast-Slow VLA Policies for Whole-Body Robotic Manipulation. Astribot Team, arXiv 2512.20188 v1, Dec 23 2025. Coverage: - Architecture: π0-FAST built on 3B PaliGemma backbone as slow pathway (RVQ-VAE action tokens, autoregressive @ 3.27 Hz) + Transformer flow-matching diffusion policy as fast pathway (32-step whole-body action chunks @ 32.3 Hz on RTX 4090) - Slow-fast interface: written-and-read latent representation buffer storing instruction embeddings + learnable fusion- query embeddings + (optional) reasoning/action tokens. Refreshed at 1-3 Hz by slow side, consumed at 25-30 Hz by fast side - Whole-body action tokenizer: residual-VQ-VAE with three parallel stream-specific codebooks (position / SO(3) rotation / gripper), each 1024 entries, over 29-d whole-body action on Astribot S1's 25 DoF (length-36 tokens vs FAST's 81-205) - Joint end-to-end training. Stage 1 = 30 epochs on 24x H100, Stage 2 = 12 epochs (wall-clock not reported) - Real-world popcorn-scooping task (20 ID + 10 OOD trials) + beverage cabinet side task. No LIBERO / SimplerEnv / RoboTwin Comparative analysis (~200 lines) along 5 async-VLA dimensions: - Intra-policy parallel (DuoCore-FS) vs intra-stack shared params (Fast-in-Slow) vs inter-system cloud-edge (AsyncVLA / Hirose+Levine) vs intra-model chunk-boundary (RTC / PI+Berkeley) vs inter-frame subgoal-image (pi0.7 BAGEL-14B 4s refresh) - 12-axis side-by-side table vs AsyncVLA / Fast-in-Slow / RTC / pi0.7 / GR00T N1.x / Helix-02 / pi0.5+Hi-Robot - Closest peer: Helix (Figure) for the parallel slow-fast pattern; Hume for action-aware slow side; Fast-in-Slow for joint training. DuoCore-FS combines all three plus end-to-end joint training - System 0/1/2 placement: explicit S1/S2, no System 0 (no balance prior, no tactile fingertip layer) Updates to Home.md, Review-VLA-Architecture.md (sec3 paper table under Cat F+H + sec5 Category F exemplars list), Review-System- 0-1-2.md (new row in sec6 dual-system table). Honest gaps flagged: closed implementation ("provided to commercial users by Astribot"); no weights, no code, no public reproduction path; total trainable params not stated; only one real-world task with 30 total trials; no buffer / asynchrony / joint-training / VLM-swap / slow-rate / delay-distribution ablations; no head-to-heads against FiS-VLA / Hume / OpenHelix / pi0.5-0.7. Mermaid scan after fix patterns from prior rounds: 0 issues across all 4 modified files.