The article explains how Vision-Language-Action (VLA) models enable robots to act on visual and language inputs, but argues that vision cannot capture contact dynamics like slip, force, or stability, making tactile sensing a critical complement for precise manipulation tasks.
VLAct introduces a representation-centric continued pre-training framework for Vision-Language-Action models, achieving 92.5% on RoboTwin 2.0 and surpassing all World Action Models on RoboDojo using only 16 GPUs and open data; with 20% downstream data it outperforms GR00T N1.6 on unseen GR-1 robot.
UniSteer introduces a noise‑inversion interface that lets human corrections directly train a lightweight noise actor, enabling a Vision‑Language‑Action model to improve real‑world task success from 20% to 90% within 66 minutes and outperforming DSRL and DAgger baselines.
UniSteer introduces a noise‑steering interface that lets human corrections and reinforcement learning jointly update a lightweight noise actor, enabling a Vision‑Language‑Action robot to raise task success from 20% to 90% within 66 minutes while using only two full human demonstrations.
The World Critic Model (WCM) augments the critic in vision‑language‑action reinforcement learning with future state prediction, enabling robots to evaluate not only the current value but also anticipate upcoming dynamics, which dramatically improves both in‑distribution and out‑of‑distribution performance across multiple benchmarks.
The ECCV‑2026 paper introduces DriveTeach‑VLA, a vision‑language‑action model that improves autonomous driving by distilling traffic‑aware visual cues and projecting BEV trajectories onto image pixels, achieving state‑of‑the‑art PDMS scores of 90.4 on NAVSIM and up to 92.7 with a trajectory selector, while detailing the training pipeline, visual distillation, 2D‑TGP prompting, and extensive ablations.
The paper analyzes why Vision‑Language‑Action models fail when task instructions are paraphrased, demonstrates that language semantics remain partially encoded, and shows that the Grounded Semantic Re‑Binding (GSR) redesign of the language‑to‑action information flow improves instruction generalization by up to 40% across multiple VLA architectures.
The paper analyzes why Vision‑Language‑Action models fail when instructions are paraphrased, designs a grounded semantic re‑binding (GSR) method that restructures language flow, and demonstrates 20‑40% gains in instruction generalization across several VLA architectures on the LIBERO‑Para benchmark.
TurboVLA, a real‑time vision‑language‑action model from Huazhong University of Science and Technology and Huawei, bypasses the large language model bottleneck by directly fusing visual and language features, achieving 32 Hz online action prediction on a single RTX 4090 with only 0.2 B parameters and 0.9 GB VRAM, while maintaining high success rates across LIBERO, RoboTwin 2.0, and real‑robot tasks.
The article examines why vision‑language‑action models alone cannot fully understand physical environments, outlines the limitations of visual‑only embodied AI, and surveys recent algorithmic efforts—such as VTLA, N0‑VTLA, OmniVTLA—to embed tactile perception as a “touch‑neuron” for finer physical interaction.
The article reviews the Vision‑Language‑Action (VLA) safety landscape, classifies attacks and defenses across training and inference phases, highlights the multimodal attack surface, real‑time constraints, and simulation‑to‑reality gaps, and proposes a fast‑slow dual‑loop defense architecture for safe embodied AI.
The paper presents Jetson-PI, an open‑source VLA real‑time control framework for low‑power edge devices that tackles inference latency and perception‑action misalignment through foresight‑aligned asynchronous correction, confidence‑based scheduling, and edge‑engine optimizations, raising Jetson Orin control frequency from 0.7 Hz to 6.06 Hz and improving task success rates on LIBERO benchmarks and a real‑world clothing‑folding robot.
Harness VLA introduces a Harness Layer that orchestrates frozen Vision‑Language‑Action models with an Agentic Planner, dramatically improving generalization on challenging robot benchmarks—achieving 82.4% success on LIBERO‑Pro versus 18.2% for NVIDIA Cap‑X—while remaining model‑agnostic and open‑source.
The talk presents Pelican-VLA 0.5, a unified Vision‑Language‑Action model that leverages attention‑level generalization without task‑specific supervision, achieving over 91% success on RoboTwin benchmarks and demonstrating early zero‑shot generalization through a novel Reasoning Slots bottleneck.
ATHENA introduces a data‑curation framework for billion‑parameter multi‑task Vision‑Language‑Action models that extends influence functions via Kronecker gradient compression and a multitask influence interaction scheme, achieving a 313× reduction in compute (from 8054.6 to 25.7 GPU‑hours) and improving task success rates while using fewer, higher‑value demonstrations.
LabVLA introduces a Vision‑Language‑Action paradigm and a knowledge‑enhanced simulation engine to teach AI systems how to plan and execute real‑world scientific experiments, achieving 71.1%/70.0% success in simulated benchmarks and demonstrating comparable performance on a real Franka robot while highlighting remaining challenges for fully autonomous lab assistants.
OpenHLM presents an open‑source VLA recipe that lets humanoid robots coordinate arms, torso, legs, and feet under vision‑language commands, using decoupled whole‑body teleoperation, multi‑step flow generation, and low‑cost data sources to achieve superior whole‑body loco‑manipulation performance on the HLM‑12 benchmark.
LabVLA introduces a vision‑language‑action framework and a knowledge‑enhanced simulation engine to enable AI models to learn and generalize scientific lab manipulation, achieving 71% success on benchmark tasks and demonstrating real‑world performance on a Franka robot, while outlining current limitations and future directions.
The article analyzes VLA‑JEPA, a JEPA‑style pre‑training framework that combines limited robot trajectories with abundant human video to build a latent world model for Vision‑Language‑Action tasks, showing improved robustness and high success rates across simulated and real‑robot benchmarks.
MINT introduces a spectrally disentangled tokenization and intent‑driven strategy that lets Vision‑Language‑Action models generalize compositionally, transfer with a single demonstration, and achieve state‑of‑the‑art performance and robustness across benchmark suites and real‑world robot experiments.
The paper introduces MetaFine, a diagnostic meta‑evaluation framework that moves robot manipulation assessment from a simple success/failure binary to a three‑dimensional analysis of understanding, perception, and behavior, revealing up to 70% over‑estimation in traditional benchmarks and offering a hybrid real‑sim testing pipeline for fair, reproducible results.
LoongForge applies system‑level optimizations—async data prefetch, fine‑grained communication‑compute overlap via a Megatron distributed optimizer, and per‑microbatch CUDA Graph scheduling—to the GR00T N1.6 Vision‑Language‑Action model, delivering up to 2.3× higher training throughput and a 56.6% reduction in overall training time on an 8×A800 cluster.
AutoMoT introduces a unified Vision‑Language‑Action model that combines a 4B Qwen3‑VL understanding expert with a 1.6B action expert via layer‑wise shared attention and asynchronous inference, achieving state‑of‑the‑art results on Bench2Drive and nuScenes while preserving general VLM capabilities.
The article reviews the emerging World Action Model (WAM) paradigm, critiques the limitations of Vision‑Language‑Action models, outlines cascaded and joint WAM architectures, discusses required data sources, evaluation metrics, and future challenges, positioning WAM as a new foundational approach for embodied AI.
HiF-VLA introduces a motion‑centric bidirectional spatiotemporal reasoning framework with a joint‑expert module that simultaneously predicts future visual motion and generates high‑precision action sequences, eliminating visual redundancy, cutting inference latency and memory usage, and achieving superior success rates on long‑horizon benchmarks such as CALVIN and LIBERO‑LONG.
This article surveys recent advances in robot world models, explaining why predictive models are essential for embodied intelligence, how they integrate with Vision‑Language‑Action systems, the various architectural approaches, benchmark trends, and the remaining challenges for reliable deployment.
LARYBench (Latent Action Representation Yielding Benchmark) provides the first systematic, ImageNet‑scale evaluation for implicit action representations derived from large‑scale human video, decoupling representation quality from downstream control, and shows that general‑purpose vision models outperform specialized embodied models in both action generalization and control precision across diverse robot morphologies and environments.
StreamingVLA introduces action‑flow matching and adaptive early observation to parallelize generation, execution, and perception in vision‑language‑action models, cutting per‑action latency from 49.9 ms to 31.6 ms, reducing stall time 6.5‑fold, and achieving up to 2.4× end‑to‑end speedup in LIBERO benchmarks and real‑world robot tests.
Generalist AI’s GEN-1 model achieves over 99% success, 2‑3× speed gains with only a tenth of the data, and its founders argue that vision‑language‑action (VLA) models are merely a crutch, urging a shift toward goal‑driven, fully‑scratch training for physical AGI.
The Point‑VLA method introduced by Qianxun AI’s Gaoyang team tackles the fundamental limits of language‑only instruction in vision‑language‑action models by adding visual grounding via bounding‑box cues, boosting real‑robot success rates from 32.4% to 92.5% across six challenging tasks.
This roundup highlights six recent embodied AI papers that advance world models and vision‑language‑action (VLA) techniques, covering DreamDojo's massive first‑person video model, LingBot‑World simulator, Agent World Model generator, BagelVLA, ACoT‑VLA, and the closed‑loop World‑VLA‑Loop framework.
This roundup presents five recent AI papers covering hierarchical sparse attention for ultra‑long context, Nvidia's Alpamayo‑R1 VLA model for autonomous driving, the non‑autoregressive F5‑TTS system, LatentMAS for latent‑space multi‑agent collaboration, and Deeper‑GXX that deepens arbitrary graph neural networks, highlighting each method's key innovations and reported performance gains.
The paper introduces RoboMonkey, a framework that applies a generate‑and‑verify paradigm and test‑time scaling to Vision‑Language‑Action models, showing that increasing sampling and verification at inference dramatically reduces action error across multiple VLA architectures, and presents scalable verifier training, synthetic data augmentation, and efficient deployment strategies.
World-Env introduces a virtual training sandbox that eliminates physical interaction, dramatically improves data efficiency with just five expert demos per task, and employs a vision‑language model as a semantic judge to dynamically terminate actions, enabling safe, high‑performing VLA post‑training across diverse robotic benchmarks.
Helix is a groundbreaking Vision‑Language‑Action model that integrates perception, language understanding, and motor control, enabling humanoid robots to perform full upper‑body continuous movements, collaborate across multiple robots, grasp any household object via natural language, and run on low‑power embedded GPUs for commercial use.
