A living research wiki for Vision-Language-Action models · robot manipulation · humanoid intelligence · tactile & force interaction · world models.

📄 396 pages	🗓 8 venues	📖 16+ cross-paper reviews	🔄 updated 2026-07-25

New here? Start with one of these: 🧭 VLA Architectures — choose the model family & action decoder · 📈 π Series Evolution — the production-VLA baseline lineage · 🏆 IROS 2026 Survey 🆕 · RSS 2026 Survey — the latest venue surveys · 🆕 Latest Papers — preprint tracker (pre-publication reviews) · 🕸️ Knowledge Graph — the visual map

---

Pick your question → open the fold-out for a 4-line orientation → follow the 📄 deep-dive link for the full survey (trend arc · approach taxonomy · limitations). Every deep-dive page carries a matching "State of the Field" section, updated Aug 2026.

Which VLA architecture should we use? — flow matching dominates; ordered tokens revived AR

📄 Deep dive → VLA Architectures

📈 Trend	AR tokens (2024) → flow-matching experts standard (2025) → flow universal in flagships + AR revived via OAT's anytime prefix decoding (2026)
⚖️ Approaches	AR discrete tokens (LLM-native / lossy, slow) · flow-matching expert (precise, production-proven / no native log-probs) · discrete diffusion in-VLM (parallel decode / little traction)
⚠️ Open	No matched-scale comparison across the three families; latency rarely published

How should the VLM connect to the action expert? — evidence tilts to cross-attention, margins small

📄 Deep dive → VLM↔Action Connection

📈 Trend	2025 four-way standoff → first direct ablations in 2026: cross-attention wins in RobotManip (87.5 vs 87.0) and MM-DiT beats naive DiT in Ψ₀
⚖️ Approaches	same-stack MoE+prefix-KV (tight / hard to retrofit) · cross-attention (decoupled, small experts / one-way) · concatenation (simplest / re-attends per step) · adaLN token (cheapest / bottleneck)
⚠️ Open	~0.5 pp margins on one benchmark; the two Qwen flagships disagree internally

How should attention / KV-cache be designed? — the runtime now drives the choice

📄 Deep dive → VLA Attention

📈 Trend	Design follows the runtime since 2025 (RTC/streaming/prefix-KV); 2026 adds linear-attention backbones (Qwen3.5) with zero control-quality ablation
⚖️ Approaches	prefix-KV + expert branch (cache reuse) · full re-attention per chunk (freshness) · streaming-token designs
⚠️ Open	Latency is the field's least-reported number; multi-view KV behavior is folklore

Should vision be built outside the VLM? — the ViT is the proven bottleneck; fix = action supervision

📄 Deep dive → Independent Visual Representation

📈 Trend	VLM4VLA: VLM scores don't predict control; frozen ViT −21~42 pp; action-supervised vision FT +18.1 pp — the one proven fix. 2026 splits into geometry add-ons vs dynamics encoders
⚖️ Approaches	action-align the VLM ViT (cheap, proven) · bolt-on metric 3D (spatial gains / calibration tax) · video-dynamics encoders (physics priors / loses VLM semantics)
⚠️ Open	Effects visible only under OOD stress; no standardized recipe

Which training framework / codebase? — modular vs pipeline; RL support is the new differentiator

📄 Deep dive → VLA Training Frameworks

📈 Trend	StarVLA (modular) vs TRI VLA Foundry (pipeline) → RSS 2026 adds the RL tier (RLux-VLA)
⚠️ Open	Nothing covers pretrain→SFT→RL→evaluation end-to-end; flagship recipes irreproducible in public stacks

How do we make inference real-time? — from test-time patches to trained-in continuation

📄 Deep dive → Real-Time Execution

📈 Trend	RTC (test-time, 2025) → training-time RTC → Legato native continuation (beats RTC ~10%) + OAT anytime decoding (2026)
⚖️ Approaches	test-time guidance (drop-in / unstable on some models) · trained continuation (smooth / retrain) · anytime tokens (compute dial / AR-only) · async dual-rate (decoupled rates / complexity)
⚠️ Open	ms/chunk almost never published; robustness features raise the denoise-step bill (4→10)

How does a deployed policy improve from experience? 🆕 — production-proven at RSS 2026

📄 Deep dive → RL for VLA

📈 Trend	"Impractical" (2025) → log-prob wall cracked 3 ways → π*0.6/RECAP in production: 2× throughput, ~½ failures on laundry/boxes/espresso
⚖️ Approaches	advantage conditioning (no log-probs, eats corrections / coarse credit) · direct PG on flow (principled / machinery) · residual RL (safe / capped) · world-model RFT (no rollouts / WM fidelity)
⚠️ Open	Loops are domain-narrow; exploration safety procedural; forgetting is milder than feared (ICML Oral: pretrained VLAs resist it) but untested under repeated RL

How do we handle long-horizon tasks? — from bigger context to agentic decomposition

📄 Deep dive → VLA Memory · System 0/1/2

📈 Trend	In-policy memory modules (2025) → selective key-frame history + agentic planner/executor with notebook memory (RobotNav: EQA SOTA, −77% steps)
⚖️ Approaches	in-policy memory (fast / shortcut risk) · episodic retrieval (scales / retrieval-bound) · agentic notebook (auditable / planner latency)
⚠️ Open	Manipulation autonomy record is 11 stages / 2.5 min; no agentic-manipulation demo yet

Do world models help VLA? — no longer hypothetical: dynamics pre-training beats VLAs on hardware

📄 Deep dive → World Models · WAM vs VLA

📈 Trend	Serving role (2025) → challenger evidence at RSS 2026: LDA-1B +48% dexterous over π0.5; mimic-video 10× sample efficiency
⚖️ Approaches	data engine (mature / fidelity ceiling) · evaluator (promising / action-input gap — ICML's dWorldEval is the first crack) · VAM backbone (dynamics priors / loses VLM depth) · unified WM+policy (all data tiers / heaviest) · aux losses (free / weak)
⚠️ Open	Contact physics fidelity; VAM claims await matched-scale replication; 20B video models price out labs

Can human video substitute for robot data? 🆕 — yes; the field split on how

📄 Deep dive → Human Video → Robot Transfer

📈 Trend	Retargeting pipelines (2024–25) → three camps at RSS 2026: emergence (PI: co-train, ~2× above a diversity threshold) vs decoupling (Ψ₀: 800 h + 30 h beats 10× corpora) vs synthesis (RobotManip: 24.8k h H2R)
⚖️ Approaches	co-train (no pipeline / threshold, gripper-only evidence) · staged decouple (data-efficient, humanoid-proven / 2-stage) · synthesize (unlimited scale / artifact ceiling)
⚠️ Open	Nobody has run both recipes on the same platform — the field's most valuable missing experiment

What data should we co-train on? 🆕 — the most settled question on this map

📄 Deep dive → LBM Co-training (+ RSS 89-policy sequel)

📈 Trend	Folklore → measurement: two LBM studies (89 policies, 58k+2,835 rollouts) + independent Qwen/VLM4VLA confirmations
✅ Verdicts	VL + cross-embodiment data help cumulatively · discrete robot-action tokens don't (3× replicated; latent-action tokens as VLM supervision do — ICML Oral) · co-training must be joint, not sequential
⚠️ Open	Mixture ratios are art (λ=0.1 vs 1.0, undiscussed); no per-sample attribution

How do we evaluate policies credibly? 🆕 — the in-distribution era is ending

📄 Deep dive → VLA Evaluation

📈 Trend	Triple indictment (from-scratch ≈ pretrained in-distribution; LIBERO-X 39.4→8.2 collapse; weak sim-real correlation) → RSS 2026 infrastructure: PolaRiS real-to-sim with validated rank correlation
⚖️ Approaches	static sim (cheap / saturated) · perturbation pyramids (diagnostic / still sim) · real-to-sim (reality-anchored / contact physics) · rigorous real stats (ground truth / cost)
⚠️ Open	Contact-rich sim evaluation missing everywhere; OOD reporting still voluntary

How do we improve dexterous manipulation? — touch moved from observation to prediction

📄 Deep dive → Dexterous Manipulation · Tactile VLA

📈 Trend	RL still owns in-hand skills; IL changed paradigm at RSS 2026 — ViTacFormer forecasts future contact (+50%, 11-stage record); CGP commands contacts, not poses
⚖️ Approaches	sim-to-real RL (reactive / per-skill, tactile sim gap) · predictive visuo-tactile IL (breadth / rig cost) · dexterous VLAs (language / trails specialists)
⚠️ Open	No dexterous foundation model; precision insertion unsolved everywhere (Ψ₀ 2/10, screws 0/10)

How do we generalize across robots? — alignment is the precondition for data scaling

📄 Deep dive → Cross-Embodiment

📈 Trend	Question flipped in 2026: RobotManip showed misaligned action spaces produce no scaling law; camera-frame EEF scales + transfers zero-shot. RSS extended to hands (DexGrasp-Zero 85% zero-shot, One-Hand 81.9%)
⚖️ Approaches	canonical padded tensors (simple, proven / curated slots) · camera-frame delta EEF (best transfer / calibration tax) · morphology graphs/URDFs (anatomy-grounded / hands-only) · prompts + history (no arch change / weak alone)
⚠️ Open	Joint-space zero-shot <5%; Franka-class morphology gaps resist; gripper↔hand untried

How do we run a VLA on a humanoid? — the triple-system recipe became the open reference

📄 Deep dive → Humanoid VLA · System 0/1/2

📈 Trend	RSS 2026 = humanoid loco-manipulation's coming-out; Ψ₀'s VLM + MM-DiT + RL-lower-body triple system beats 10×-data baselines by +40 pp
⚖️ Approaches	whole-body end-to-end (expressive / unstable, data-hungry) · triple-system (stable, data-efficient / agility capped) · teleop rigs vs robot-free human interfaces (quality vs scale)
⚠️ Open	Per-task fine-tuning in every loop; no humanoid OOD protocol — evaluation lags arms by a generation

---

Newest first — full history in Changelog. Pre-publication preprint reviews live in Latest Papers.

Date	Page	What it adds
07-25	RSS 2026 survey + 15 paper pages	Sydney; 210 papers, ~116 in scope; π*0.6/RECAP flagship; figure-illustrated reviews
07-25	Ψ₀ (in-depth)	Open humanoid foundation model — 800 h human video + 30 h robot data beats 10× corpora
07-24	Qwen Team's VLA Program	Cross-paper: VLM4VLA → Qwen-VLA → Qwen-Robot Suite (5 in-depth reviews)
07-24	Qwen-RobotManip · RobotNav · RobotWorld	The Qwen-Robot Suite, deep-read
06-11	VLA Training Frameworks	StarVLA vs TRI VLA Foundry
06-10	RoboMME (in-depth)	Memory-implementation breakdown

---

→ Topic reviews catalog — cross-paper topic reviews, lab programs, latest-paper reviews · Per-paper long-forms — single-paper deep-dives.

	Most-used entries
Topic reviews	VLA Architectures · VLA Hybrid Architectures 🆕 · Multi-Task VLA 🆕 · VLM↔Action · RL for VLA · World Models · Dexterous · Dex-Hand Data Pyramid 🆕 · Tactile · Cross-Embodiment · Single-Checkpoint Multi-Robot 🆕 · Egocentric Video Pre-Training 🆕 · Humanoid · Memory · In-Context Imitation 🆕
Lab programs	π series · GR00T N1→N1.7 · RoboTTT (context scaling) 🆕 · Qwen VLA program · Ψ₀ · DreamZero 🆕
ML foundations	ML hub · Attention Variants · Normalization

---

Venue	Year	Survey / Index	Character
RSS	2026	Survey 🆕	Method frontier — the improvement loop, humanoids, hands, evaluation (210 papers)
ICRA	2026	Survey	Deployment & sensors at scale (728 in-scope papers, 12 topic pages)
ICML	2026	Index	99 manipulation papers + 🏅 Top-20
CVPR	2026	Survey	Vision-first VLA / perception (also 2025)
ICLR	2026	Survey	Architecture & representation (~211 papers)
NeurIPS	2025	Survey	Foundation-model training methods
CoRL	2026	Survey 🆕	Robot learning (687 accepted; preliminary — official list pending)
CoRL	2025	Survey	Robot learning / benchmarks
IROS	2026	Survey 🆕	Robustness/efficiency/deployment of learned & VLA policies (1,900+ papers; titles-level, pre-conf)
IROS	2025	Survey	Hardware, systems, deployment

---

OpenVLA (open 7B AR baseline) · ReKep (training-free keypoint constraints) · AgiBot World Colosseo (1M+ trajectory dataset) · RoboBrain 2.0 (embodied-reasoning VLM)

---

_{🛠 Contributing: page templates & house rules → Maintenance. Summaries are compiled from public sources and are not substitutes for the original papers — verify numbers before citing.}