| ici-vla-in-context-imitation-spatiotemporally-aligned-demonstrations ICI-VLA: In-Context Imitation with Spatiotemporally Aligned Demonstrations for Vision-Language-Action Models | arXiv / IEEE CDC 2026 | ⭐ HIGH PRIORITY: Frozen-weight in-context task adaptation via aligned demo retrieval, beating naive ICL and OpenVLA-OFT on LIBERO/RoboTwin/real robots. |
| infinova-infinite-novel-view-augmentation-viewpoint-invariant-robot-policies InfiNoVA: Infinite Novel View Augmentation for Viewpoint Invariant Robot Policies | arXiv / TPC Workshop @ SC’26 | ⭐ HIGH PRIORITY: Gaussian-splat-based viewpoint augmentation gives 5.4x robustness to unseen camera placements over prior VISTA baseline. |
| turbovla-real-time-vla-32hz-rtx4090 TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM | arXiv | 0.2B-param VLA hitting 97.7% LIBERO at 31ms/<1GB VRAM on a consumer GPU, bypassing the usual LLM-backbone bottleneck. |
| devol-one-autoregressive-mixture-transformers-vla-latent-world-modeling Devol-ONE: One Autoregressive Mixture of Transformers to Unify Vision-Language-Action and Latent World Modeling | arXiv | Single MoT architecture unifying VLA control and V-JEPA-initialized world modeling; 98.4% LIBERO, 71.4% zero-shot LIBERO-Plus. |
| where-predictive-supervision-goes-shapes-what-vla-policies-learn Where Predictive Supervision Goes Shapes What VLA Policies Learn | arXiv | Analysis paper arguing how future-prediction supervision enters a VLA’s visual stream, not just its presence, drives robustness. |
| ati-vla-action-centric-predictive-vla-actionable-alignment-adaptive-injection ATI-VLA: Action-Centric Predictive Vision-Language-Action Models via Actionable Alignment Then Adaptive Injection | arXiv / NeurIPS 2026 | Fixes known modality-misalignment weakness of predictive VLAs; 97.9% LIBERO, 70% real-robot long-horizon. |
| fewer-tokens-better-action-gpt-6-astra-robot-agents Fewer Tokens, Better Action: GPT-6 Astra Robot Agents with 14% Higher Success Rate but 65% Fewer Tokens | arXiv | Code-execution planning framework (PyRUA-Lean) cuts LLM-agent token/cost overhead while raising task success. |
| wbag-whole-body-attached-geometry-safety-framework-vla WBAG: A Whole-Body and Attached-Geometry Safety Framework for Vision-Language-Action Manipulation | arXiv | Safety layer covering the whole robot body and grasped-object geometry, not just the end-effector; best on SafeLIBERO. |
| is-success-all-you-need-input-perturbations-vla-tabletop-manipulation Is Success All You Need? Investigating the Impact of Input Perturbations on VLA Behaviour in Tabletop Manipulation Tasks | arXiv | Shows success-rate alone hides behavioral degradation under perturbation; proposes trajectory-quality metrics. |
| chunkvla-am-parallel-action-chunking-additive-manufacturing ChunkVLA-AM: Parallel Action Chunking for Vision-Language-Action Robot Control in Additive Manufacturing | arXiv | Real-hardware case study deploying action-chunked OpenVLA-OFT for 3D-print post-processing (92.9% over 42 trials). |
| continuous-conditioning-vlas-emg-visual-task-descriptors Continuous Conditioning of VLAs with Augmenting EMG and Visual Task Descriptors | arXiv / IROS WORLDS Workshop 2026 | Shows EMG and visual-annotation conditioning can beat language prompts for VLA task specification in cluttered scenes. |
| encoded-but-not-in-control-grounding-gap-vision-language-robot-policies Encoded but Not in Control: Revealing the Grounding Gap in Vision-Language Robot Policies | arXiv | Finds target info is encoded internally but not acted upon — a representational-to-behavioral “grounding gap” in VLAs/WAMs. |
| a-safe-action-is-not-enough-feasible-future-decoding-vla-policies A Safe Action Is Not Enough: Feasible-Future Decoding for Vision-Language-Action Policies | arXiv | Introduces a “feasibility-likelihood gap” and a training-free reranker (VICS-G) for safer VLA decoding. |
| smart-zero-shot-sim-to-real-articulated-object-manipulation-synthetic-pretraining SMART: Zero-Shot Sim-to-Real Articulated Object Manipulation via Large-Scale Synthetic Pretraining | arXiv | ⭐ HIGH PRIORITY: 1M+ synthetic demo dataset (2,500+ articulated objects) enabling 76.7% zero-shot real-robot success. |
| vla-acl-action-consistent-visual-token-pruning VLA-ACL: Action-Consistent Visual Token Pruning for Efficient Vision-Language-Action Models | arXiv | Action-consistency-supervised pruning cuts 87.5% of visual tokens, 75% compute, 1.5x speedup on frozen VLAs. |
| esp-energy-score-policy-one-step-multimodal-action-generation ESP: Energy-Score Policy for One-Step Multimodal Action Generation | arXiv | Energy-score training objective enables one-pass multimodal action generation, cutting latency 28.4ms→3.35ms. |
| stairvla-stage-aware-hierarchical-action-generation StairVLA: Stage-Aware Hierarchical Action Generation for Vision-Language-Action Models | arXiv | Stage-aware high-level/refiner split improves both LIBERO accuracy (96.5%→97.8%) and latency (115ms→44ms). |
| bigym-2-benchmarking-humanoid-household-manipulation BiGym 2.0: Benchmarking Learned and Agent-Developed Policies for Humanoid Household Manipulation | arXiv | New open-sourced Unitree G1 humanoid benchmark comparing VLA fine-tuning, IL, demo-RL, and LLM coding agents head-to-head. |
| vidal-visual-dynamics-grounded-action-latent-space-vla ViDAL: A Visual Dynamics-Grounded Action Latent Space for Vision-Language-Action Models | arXiv | Grounds latent action space in predicted visual dynamics; 98.1% LIBERO plus two real-robot platforms. |
| warefly-vla-uav-navigation-human-tracking-smart-warehouses WareFly-VLA: A Vision-Language-Action Framework for UAV Navigation and Human Tracking in Smart Warehouses | arXiv (submitted to Robotics and Autonomous Systems) | Rare VLA benchmark for drones rather than arms/humanoids; exposes a sharp generalization drop for repurposed manipulation VLAs. |
| acttune-action-aware-precision-gpu-adaptation-energy-efficient-vla-inference ActTune: Action-Aware Precision and GPU Operating-Point Adaptation for Energy-Efficient Vision-Language-Action Inference | arXiv | Per-call precision/GPU-tuning cuts VLA inference energy up to 76.8% with 2.02x speedup and higher success. |
| mim-vla-physical-interaction-representations-gripper-motor-feedback MIM-VLA: Learning Physical Interaction Representations from Gripper Motor Feedback | arXiv | Sensor-free object-stiffness discrimination (75% vs 48.8% baseline) from gripper motor telemetry alone. |
| vommi-collecting-leveraging-portable-demonstrations-mobile-manipulation VOMMI: Collecting and Leveraging Portable Demonstrations for Mobile Manipulation | arXiv | ⭐ HIGH PRIORITY: $100-200 portable RGB rig for mobile-manipulation data collection without the robot, beating OpenPI 0.5 by 8.3pts. |
| adapting-vla-models-unknown-visual-disruptions-during-execution Adapting Vision-Language-Action Models to Unknown Visual Disruptions During Execution | arXiv | ⭐ HIGH PRIORITY: SALT uses self-supervised test-time adaptation from the unexecuted action-chunk remainder, +9pts success under visual corruption. |
| commit-while-futures-agree-consequence-aware-adaptive-action-chunking Commit While Futures Agree: Consequence-Aware Adaptive Action Chunking for Robot Manipulation | arXiv | World-model-guided adaptive chunk-commit length cuts failure rate up to 71.8% relative, training-free. |
| towards-efficient-robotic-manipulation-models-self-recursive-pruning Towards Efficient Robotic Manipulation Models with Self-Recursive Pruning | arXiv | Training-free recursive pruning retains >90% of OpenVLA’s success at 50% weight pruning on LIBERO-Object. |