Add CVPR 2026 manipulation survey + 29 per-paper pages
Retrospective survey of the CVPR 2026 manipulation cluster (Denver,
June 3-7, 2026). 16,092 -> 4,090 accepted (25.4%). Embodied-AI /
robotics share of accepted papers grew from ~2.9% (CVPR 2025) to
~6.2%.
Survey landing: CVPR-2026-VLA-Manipulation-Survey.md — 10 buckets,
5 distinguishing trends vs. CVPR 2025, full CVPR 2025 -> CoRL 2025 ->
NeurIPS 2025 -> ICLR 2026 -> CVPR 2026 lineage map, validation notes
on common miscitations.
Year landing: CVPR-2026.md — quick paper index by category, trends,
relevant workshops.
29 per-paper pages with mermaid diagram + Problem / Method / Results
/ Significance / Links sections (ICLR-2026 template):
VLA architecture
- OptimusVLA, Boost-FAN, Fast-ThinkAct (NVIDIA Taiwan, -89.3% latency),
DM0, RDT2 (THU-ML, 7B + RVQ), DiT4DiT (98.6% LIBERO)
Reasoning / CoT / self-correction
- ACoT-VLA (AgiBot, action-space CoT), Action-Sketcher (Highlight),
CycleVLA (Oxford+Cambridge, MBR decoding)
World-model / RL
- GigaBrain-0.5M (RAMP +~30% over RECAP), Robo-Dopamine (BAAI PRM),
CoWVLA (latent-motion subgoals)
Egocentric & human-video
- UniDex (8 dex hands), EgoVLA (UCSD+UIUC+MIT+NVIDIA), EgoScale (NVIDIA
GEAR, 20kh — the GR00T N1.7 data delta), FunREC (ETH+MPI functional
3D scenes)
Humanoid loco-manipulation
- VIRAL (NVIDIA+Berkeley+CMU+CUHK, 54 G1 cycles), Open-Sim-to-Real
(companion, +31.7% over teleop), Humanoid-GPT (2B-frame mocap GPT),
Gallant (voxel humanoid locomotion)
Affordance / grounding
- RealVLG-R1 (Tongji, 11B grounding benchmark), SaPaVe (PKU+BAAI
active perception), PanoAffordanceNet (likely)
Diffusion / mobile / tactile / bimanual
- AnchorVLA (likely), EchoVLA (likely), HapticVLA (likely), HandX
Benchmarks / robustness
- LIBERO-Plus (95% -> <30% under perturbation — robustness wake-up
call), RC-NF (sub-100ms anomaly detection)
Wired into Home.md (Latest surveys + What's new + by-conference
table), _Sidebar.md (new CVPR 2026 section with 8 sub-buckets),
CVPR.md (2026 row prepended).