ICLR 2026

ICLR 2026 received 19,525 valid submissions and accepted 5,355 papers (225 orals; 27.4% acceptance — a three-year low). 779 were desk-rejected and 5,042 withdrawn, so 13,763 submissions received an accept/reject decision (8,408 rejected). VLA / robot manipulation is one of the largest growth areas — ~210 robot- or VLA-related accepted papers by keyword count, vs. <100 in 2025.

This index was rebuilt against the official accepted list at https://iclr.cc/virtual/2026/papers.html (data file: /static/virtual/data/iclr-2026-orals-posters.json, 2026-05-09 snapshot). Items listed under "Confirmed at ICLR 2026" have been individually verified against the OpenReview decision; companion technical reports and arXiv-only preprints that were previously cross-listed are now relocated to a separate section.

Surveys hosted in this wiki

• VLA & Manipulation Survey — five categories, 8 trend analyses, ~37 paper summaries.

---

Confirmed at ICLR 2026 — wiki pages

These wiki pages map 1-to-1 to a paper in the official ICLR 2026 accepted list. The bullet shows the official accepted title when it differs from our shorthand.

Architecture · Diffusion / discrete-diffusion action heads

• Unified Diffusion VLA — Unified Diffusion VLA: Vision-Language-Action Model via Joint Discrete Denoising Diffusion Process

Architecture · Memory · Cross-embodiment · MoE

• HAMLET — HAMLET: Switch Your Vision-Language-Action Model into a History-Aware Policy
• MemoryVLA — MemoryVLA: Perceptual-Cognitive Memory in Vision-Language-Action Models for Robotic Manipulation
• X-VLA — X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model
• UniVLA — Unified Vision-Language-Action Model
• WholeBodyVLA — WholeBodyVLA: Towards Unified Latent VLA for Whole-body Loco-manipulation Control
• Cosmos Policy — Cosmos Policy: Fine-Tuning Video Models for Visuomotor Control and Planning
• VLM4VLA — VLM4VLA: Revisiting Vision-Language-Models in Vision-Language-Action Models

Tokenizers · Quantization · Efficiency

• FASTER — FASTer: Toward Powerful and Efficient Autoregressive Vision–Language–Action Models with Learnable Action Tokenizer and Block-wise Decoding
• AutoQVLA — QVLA: Not All Channels Are Equal in Vision-Language-Action Model's Quantization

Data · Benchmarks · Simulation

• EgoDex — EgoDex: Learning Dexterous Manipulation from Large-Scale Egocentric Video
• RoboCasa365 — RoboCasa365: A Large-Scale Simulation Framework for Training and Benchmarking Generalist Robots
• RoboArena ∞ — RobotArena ∞: Scalable Robot Benchmarking via Real-to-Sim Translation
• WorldGym — WorldGym: World Model as An Environment for Policy Evaluation

Training approaches · Co-training · Reasoning · Composition

• Actions as Language — Actions as Language: Fine-Tuning VLMs into VLAs Without Catastrophic Forgetting
• InstructVLA — Vision-Language-Action Instruction Tuning: From Understanding to Manipulation
• Hybrid Training — Hybrid Training for Vision-Language-Action Models
• Embodied-R1 — Embodied-R1: Reinforced Embodied Reasoning for General Robotic Manipulation
• Compose Your Policies — Compose Your Policies! Improving Diffusion-based or Flow-based Robot Policies via Test-time Distribution-level Composition
• Ctrl-World — Ctrl-World: A Controllable Generative World Model for Robot Manipulation

Dexterous manipulation

• DexNDM — DexNDM: Closing the Reality Gap for Dexterous In-Hand Rotation via Joint-Wise Neural Dynamics Model
• UniHM — UniHM: Unified Dexterous Hand Manipulation with Vision Language Model
• RFS — RFS: Reinforcement learning with Residual flow steering for dexterous manipulation

RL for VLA → see also RL

• PLD — Self-Improving Vision-Language-Action Models with Data Generation via Residual RL
• Stage-Aware RL — SARM: Stage-Aware Reward Modeling for Long Horizon Robot Manipulation
• VITA — VITA: Vision-to-Action Flow Matching Policy
• SimpleVLA-RL — SimpleVLA-RL: Scaling VLA Training via Reinforcement Learning

---

Confirmed at ICLR 2026 — newly added wiki pages (May 2026 batch)

These pages were created in the May 2026 wiki rebuild, one per accepted paper. Each page follows the standard template (mermaid pipeline · Problem · Method · Results · Significance · Links).

Architecture · VLA backbones

• OneTwoVLA — Unified VLA with Adaptive Reasoning (System-1/System-2 switching inside one model)
• HybridVLA — Collaborative Diffusion and Autoregression in a Unified VLA
• Vlaser — VLA with Synergistic Embodied Reasoning (Vlaser-6M reasoning corpus)
• villa-X — Enhancing Latent Action Modeling in VLA (Vision-Language-Latent-Action)
• MetaVLA — Unified Meta Co-Training for Efficient Embodied Adaptation (post-training, backbone-agnostic)
• PixelVLA — Advancing Pixel-level Understanding in VLA (multiscale pixel-aware encoder + visual prompting)
• TwinVLA — Twin Single-Arm VLAs for Bimanual Manipulation (no bimanual pre-training)
• VER — Vision Expert Transformer via Foundation Distillation and Dynamic Routing (<0.4%-param router)

Spatial / 3D grounding for VLA

• Spatial Forcing — Implicit alignment to a 3D foundation model
• Spatially Guided Training — SP-VLA two-stage spatial pretraining (SimplerEnv 66 → 84%)
• From Spatial to Actions (FALCON) — 3D tokens injected at the action head
• EquAct — SE(3)-equivariant transformer + iFiLM, 18 RLBench tasks
• PA3FF — Part-aware dense 3D feature field + Part-Aware Diffusion Policy
• Seeing Across Views — MV-RoboBench (1.7k QA, 8 subtasks)

Efficiency · Pruning · Acceleration

• SP-VLA — Joint scheduling + token pruning, 1.5×/2.4× speedup
• Action-aware Dynamic Pruning — Phase-adaptive token pruning, 1.35× on OpenVLA-OFT
• Verifier-free Test-Time Sampling — MG-Select via KL-to-masked-reference
• Block-wise Adaptive Caching for Accelerating Diffusion Policy
• Real-Time Robot Execution with Masked Action Chunking

Robustness · Adaptation

• On Robustness of VLA — RobustVLA (17 perturbations / 4 modalities, +12.6% on π₀)
• Robust Fine-tuning via Parameter Merging — pretrained ↔ finetuned weight interpolation
• Align-Then-Steer — VAE-aligned latents, +9.8% sim / +32% real cross-embodiment
• When would Vision-Proprioception Policies Fail in Robotic Manipulation?

World models · Video for actions

• WMPO — World Model-based Policy Optimization for VLA
• Genie Envisioner — Unified World Foundation Platform
• Vid2World — Video Diffusion → Interactive World Models
• ViPRA — Video Prediction for Robot Actions (+16% SIMPLER, +13% real, 22 Hz)
• Geometry-aware 4D Video Generation — 4D video gen for manipulation
• Sim2Real VLA — zero-shot synthesized-skills transfer
• Robotic Manipulation by Imitating Generated Videos Without Physical Demonstrations
• Sparse Imagination for Efficient Visual World Model Planning
• Test-Time Mixture of World Models for Embodied Agents

Embodied reasoning · Planning · CoT

• From Seeing to Doing — Bridging reasoning and decision (8 spatial benchmarks)
• OmniEVA — Embodied Versatile Planner via Task-Adaptive 3D-Grounded reasoning
• VLMgineer — VLMs as Robotic Toolsmiths
• Self-Refining VLM (ARMOR) — Robotic failure detection and reasoning
• Cortical Policy — Dual-Stream View Transformer for manipulation
• RoboInter — Holistic Intermediate Representation Suite
• Policy Contrastive Decoding for Robotic Foundation Models
• RoboOmni: Proactive Robot Manipulation in Omni-modal Context
• Self-Improving Loops for Visual Robotic Planning

Bimanual · Mobile · Dexterous

• VLBiMan — Vision-Language Anchored Bimanual from one-shot demos
• MoMaGen — Bimanual mobile manipulation demo generation under constraints
• DemoGrasp — Universal dexterous grasping from a single demonstration
• DexMove — Tactile-Guided Non-Prehensile Manipulation with dexterous hands
• D-REX — Differentiable Real-to-Sim-to-Real Engine for Dexterous Grasping
• RoboPARA — RoboPARA: Dual-Arm Robot Planning with Parallel Allocation and Recomposition Across Tasks (LLM dependency-graph dual-arm parallelism, X-DAPT dataset)
• CLAP (Coarse-to-Fine Keypoints) — Generalizable Coarse-to-Fine Robot Manipulation via Language-Aligned 3D Keypoints (VLM 3D-keypoint grounding; +12% on GemBench with 1/5 trajectories)
• SMP (MoE Diffusion Skills) — Abstracting Robot Manipulation Skills via Mixture-of-Experts Diffusion Policies (orthogonal skill basis + sticky routing)
• PF-DAG — Primary-Fine Decoupling for Action Generation in Robotic Imitation (discrete-mode + mode-conditioned MeanFlow; 56 tasks)
• OmniReset (Emergent Dexterity) — Emergent Dexterity via Diverse Resets and Large-Scale RL (programmatic diverse resets enable on-policy RL with no demos/curriculum; zero-shot real transfer)
• House of Dextra — Cross-Embodied Co-Design for Dexterous Hands (joint morphology+policy co-design; design-to-deploy a new hand in <24h)

Humanoid · Whole-body · Loco-manipulation

• BFM-Zero — Promptable Behavioral Foundation Model for Humanoid Control via unsupervised RL
• HWC-Loco — Hierarchical Whole-Body Control for robust humanoid locomotion
• From Language to Locomotion — Retargeting-free humanoid control via motion latent guidance
• LIFT (Pretrain-Finetune Humanoid) — Bridging Large-Scale Pretraining and Efficient Finetuning for Humanoid Control (off-policy SAC pretrain + physics-informed world-model finetune)
• HVD — Hierarchical Value-Decomposed Offline RL for Whole-Body Control (kinematic-tree Q-value decomposition; WB-50 dataset)

Navigation · Embodied agents

• Embodied Navigation Foundation Model (NavFoM) — 7-benchmark navigation FM
• REI-Bench — Can embodied agents understand vague instructions? (36.9% degradation finding)
• Ground Slow, Move Fast · JanusVLN · AutoFly · CompassNav · CE-Nav · From Seeing to Experiencing · Towards Physically Executable 3D Gaussian for Embodied Navigation

RL · Flow policies

• Flow Matching Policy Gradients (FPO) — RL recipe for flow-matching policies
• Translating Flow to Policy (HinFlow) — Hindsight online imitation, >2× improvement
• Guided Flow Policy — Offline RL with high-value flow guidance (144 tasks)
• Batch Online RL — What matters for self-improving robot learning from autonomous data (Q-functions + implicit extraction + diffusion)
• Learning from Constrained Demonstrators — Robot learns a better policy than its constrained (e.g. joystick) demonstrator

Data · Imitation · Benchmarks

• DataMIL — Datamodel-based selection for imitation learning
• Reflective DDVLA — Discrete Diffusion for Reflective VLA in Autonomous Driving — the only "discrete-diffusion VLA" actually at ICLR 2026 (autonomous-driving variant; the manipulation DDVLA is NeurIPS 2025 / arXiv)
• ManipEvalAgent · AutoBio · ArtVIP · Action Chunking & Exploratory Data · Demystifying Diffusion Policies · MIKASA (Memory, Benchmark & Robots) · Cross-Embodiment Offline RL · DeFI (Disentangled Fwd/Inv) · MemER (Experience Retrieval)

Sensing · Tactile

• AnyTouch 2 — General optical tactile representation for dynamic, force-aware perception (ToucHD dataset) · SpikePingpong — Spike-vision fast-slow system for high-precision robot table tennis

---

Companion preprints / technical reports — referenced in the wiki, not at ICLR 2026

These wiki pages are still useful as standalone analyses, but they describe technical reports, blog posts, or arXiv-only preprints that are not in the ICLR 2026 proceedings. Renamed from "ICLR 2026" framing.

Physical Intelligence (PI) — production reference series

• π0.6 — PI technical report (Nov 2025)
• π0.7 — PI technical report (Apr 2026)
• π*0.6 + RECAP — PI technical report (RL from experience)
• RL Tokens (RLT) — PI blog post (online RL on contact-rich tasks)
• 📊 π series evolution — π0 → π0.7 model/data/training side-by-side

Toyota Research Institute / Levine lab / Berkeley — Feb 2026 arXiv preprints

• LBM Co-training Study (TRI) — A Systematic Study of Data Modalities and Strategies for Co-training Large Behavior Models (arXiv 2602.01067, Feb 2026 — submitted post-ICLR-2026 deadline)
• AsyncVLA (Levine lab) — Hirose / Glossop / Shah / Levine, Feb 2026 arXiv
• Steerable Policies (Berkeley+PI) — Chen / Glossop / Driess / Pertsch / Levine et al., Feb 2026 arXiv

---

Wiki pages with venue still being verified

These pages exist in the wiki but the corresponding paper title was not found in the May 2026 ICLR-2026 accepted list. They may be at NeurIPS 2025, CVPR 2026, IROS 2026, or arXiv-only. Treat as preprint-quality references until confirmed:

• Discrete Diffusion VLA — only an autonomous-driving discrete-diffusion VLA was accepted; the manipulation DDVLA is likely NeurIPS 2025 / arXiv
• dVLA
• DIVA & Fast-dVLA
• HiMoE-VLA
• HyperVLA
• Human-Video Pretraining
• OmniSAT
• VLA-RFT
• XR-1

---

Methodology of this index

• Source of truth: https://iclr.cc/static/virtual/data/iclr-2026-orals-posters.json (the JSON the official virtual site uses to populate /virtual/2026/papers.html).
• Filter: keyword match on name, keywords, and topic for VLA, vision-language-action, manipulation, dexterous, humanoid, whole-body, bimanual, embodied, cross-embodiment, gripper, grasp, imitation learning, behavior cloning, flow-matching policy, diffusion policy, world model + robot/manipulation/action, etc., with negative filters for off-topic uses (image manipulation, language-model fine-tuning, etc.).
• Coverage: 211 papers identified as VLA / robot-manipulation / embodied AI / dexterous / humanoid / navigation. The "we have not yet reviewed" section above lists ~75 of the most architecturally or empirically relevant of these.
• Verification: every wiki page in this index was individually substring-matched against the JSON title set.

← Back to ICLR