Researchers from NTU, SenseTime, and Ropedia propose NEO-VLA, a native vision-language-action model that eliminates separate vision encoders by feeding images through lightweight convolutions into a multimodal backbone, allowing the action expert to directly access multi-layer visual representations for more precise robot manipulation, validated across simulations and real-world tasks.
Ropedia releases Xperience 2, a dataset combining 1,589 real-world scene scans with 4,230 first-person task recordings aligned in a common coordinate system, enabling robot learning through Real-to-Sim, action retargeting, virtual teleoperation, and long-horizon trajectory generation to establish a gold standard for physical AI data scaling.
Oxford researcher Hao Fei's 59-page paper introduces cognitive symbionts as AI-augmented human intelligence, a parallel paradigm to AGI where wearable AI systems enhance human capabilities through lifelong, embedded, real-time, private, low-friction, and proactive symbiosis, enabled by five converging technologies.
AGI vs AHIAI-Augmented Human IntelligenceCognitive Symbionts
Harbin Institute of Technology alumnus Zishen Wan joins Columbia University as assistant professor after award-winning PhD research at Georgia Tech on hardware-software co-design for embodied AI, earning top honors at DAC PhD Forum and ACM SIGDA Outstanding Dissertation Award, and now leads Wan Lab focusing on AI agent-assisted computer architecture design.
ACM SIGDA Outstanding Dissertation AwardArchEvalColumbia University
Researchers from NUS, InfRec Cardinal AI Lab, and UC Berkeley propose REA, an embodied runtime framework that enables asynchronous collaboration between high-level Agents and vision-language-action (VLA) models via a three-layer router, achieving 95.9% success on LIBERO-Pro with 25% fewer control steps and 17% less time, validated on Unitree G1 and AlohaMini2 robots.
Qoo Tech's Q0-Preview achieves 68.87% success on RoboDojo Sim using an Agent-based approach with Claude Opus 5.5, outperforming Fable 5.2 by leveraging spatial computation and a Harness framework for embodied manipulation without VLA or WAM paradigms.
Agent ArchitectureClaude Opus 5.5Dual-arm Manipulation
The Trinity framework shifts recursive self-improvement from task-level to capability-level optimization, using Creator, Explorer, and Evolver roles to co-evolve environments and harnesses across 28 multimodal benchmarks, achieving 34.1 and 26.2 point gains on seen and unseen tasks, with 60% sim-to-real transfer on robot tasks.
Q0-Preview from Qoo Tech tops the RoboDojo Sim Agent leaderboard with 68.87% success using an Agent approach powered by Claude Opus 5.5, beating Fable 5.2 without VLA or WAM by leveraging spatial computing, a Harness framework, and simulation-derived skills for dual-arm coordination and adaptive manipulation.
Anthropic releases Claude Fable 5.2 and Mythos 5.2, achieving 64.9% on RoboDojo Sim — a 37.3-point jump over its predecessor — surpassing specialized robot AI systems, while former Tesla Optimus researcher Bingjie Tang joins, signaling the 'bitter lesson' of general AI scaling extending to robotics.
Ant Group open-sources LingBot-Touch, a tactile foundation model pre-trained on 5,000 hours of real data that unifies magnetic, visual, and resistive sensor signals into a shared representation, enabling cross-sensor, cross-hand, and cross-task reuse; integration with four action models across 26 real-robot tasks raises success rates by 14–26 percentage points.
This article traces Tang Zuolin's six-year journey from HarmonyOS evangelist building Python support, community, and university programs to founding Qilinbo, an embodied AI startup on OpenHarmony that recently secured angel funding, illustrating how sustained developer advocacy bridges ecosystem gaps and enables new technology frontiers.
Touch-R1 introduces the first tactile-anchored physical reasoning model using reinforcement learning to enable cross-sensor tactile inference, correcting visual misjudgments by prioritizing tactile evidence, and achieves state-of-the-art results on the new TouchReason-Bench benchmark across nine metrics including physical property identification, ordinal reasoning, and cross-sensor consistency.
This survey from Oxford, NUS, NTU, and UC Merced presents a unified framework for video world models, distinguishing temporal coherence from world consistency, and organizing research around four operations—reading, writing, sharing, and interacting—while emphasizing that stronger claims require stronger evidence beyond visual fidelity.
A survey from CUHK MMLab and Fudan University introduces the Agentic Robots framework, organizing research across four domains—Model, Data, Environment, and Harness—with 17 functional categories, showing how agents coordinate skills, recover from failures, and turn execution experience into reusable capabilities for embodied AI.
Amap unveiled Phys AI Data at the 2026 World Internet Conference, a three-layer spatial data foundation comprising training data (embodied, spatial, synthetic), a simulation evaluation platform, and Physical AI Map, enabling robot training, validation, and deployment in real-world scenarios like malls and factories.
This review explores Physical AI as a closed-loop system integrating perception, world models, decision-making, and multi-timescale feedback, distinguishing it from mere automation by its ability to verify actions against physical consequences and adapt continuously through real-world deployment data.
Researchers from Chinese Academy of Sciences and Alibaba propose SimpleICL, a robot in-context learning framework enabling zero-shot generalization to unseen manipulation tasks by watching a single human demonstration video, using semantic discriminative data and cross-group pairing to resolve prompt ambiguity, achieving 66% success in simulation and up to 81% on real robots.
Agibot deployed six humanoid robots in a real Shanghai convenience store for a 24-hour live-streamed shift, evaluating them on actual store metrics; four passed, two failed due to engagement and stability gaps, exposing hard challenges in flexible manipulation, long-horizon planning, and dynamic obstacle avoidance, while the Genie Master framework enabled autonomous learning that lifted task success from 28.6% to 85.7% in five days.
Hong Kong University and Yisheng Technology introduce DexHoldem, a real-robot benchmark using Texas Hold'em to test both dexterous manipulation (1,470 demos, 14 skills, UR10e+Shadow hand) and embodied perception/reasoning (36 questions, 14 model combos), with metrics TCR and SPSR revealing gaps between action success and scene preservation.
Shanghai Jiao Tong University and Baidu introduce WAVE-0, a 257M-parameter latent world model that connects VLA with world action prediction, nearly doubling π₀.₅'s real-robot success rate from 38.3% to 73.3% without large-scale video pre-training, while achieving 98.7% on LIBERO and 91.47% on RoboTwin 2.0.
Tianjin University and Hefei University of Technology researchers propose FAME, a multimodal exploration framework using Conditional Flow Matching to generate multiple exploration anchors, combined with A* path planning, clustering, and hierarchical selection, achieving state-of-the-art coverage on AiMDoom and Matterport3D benchmarks with strong cross-dataset generalization.
Active 3D MappingConditional Flow MatchingCross-Dataset Generalization
Baidu Intelligent Cloud and HIT propose VLDA, a new robotic manipulation paradigm that learns compact task-relevant state changes (Delta) instead of predicting full future frames, achieving state-of-the-art results on LIBERO, LIBERO-Plus, RoboTwin 2.0 benchmarks and real-robot platforms with improved long-horizon consistency and robustness.
Harbin Institute of Technology and Baidu propose VLDA, a robotic manipulation paradigm that learns compact task-relevant state changes (Delta) instead of full future prediction, achieving 98.5% on LIBERO, 84.5% on LIBERO-Plus, and 73% real-robot success rates by using a Delta variable as an explicit transition interface between vision-language understanding and action generation.
University of Washington PhD Jianfei Duan joins NUS as Presidential Young Professor, founding MAGIC Lab; his CoRL 2026 Spotlight paper MolmoAct2 and Oral paper MolmoB0T demonstrate low-cost robots ($6K) performing embodied reasoning, building on prior work in robot learning, failure diagnosis, and generalization benchmarks.
ParallelWorld introduces test-time scaling for embodied agents by simulating multiple future action paths before execution, improving task success rates from 56% to 83% on benchmarks like ESI-Bench and LIBERO-Pro, with adaptive search width balancing performance and compute cost.
RoboCoach introduces a world-model-driven framework that actively identifies weak sub-task policies in long-horizon robot tasks, guiding targeted human demonstrations to update modular skill experts, achieving significant success rate improvements across four platforms with limited data budgets.
Harvard, NVIDIA, Berkeley, MIT, and Stanford researchers introduce RAX, a framework where a coding agent autonomously explores robot perception and control interfaces, writes executable programs from successful trials, and accumulates reusable skills—achieving 37.4% average gains over prior state-of-the-art on five benchmarks without any human demonstrations or model fine-tuning.
Harvard, NVIDIA, Berkeley, MIT, and Stanford researchers introduce RAX, a framework that scales online exploration for robot agents, achieving 37.4% average improvement over state-of-the-art on five benchmarks without any human demonstrations, using frozen model parameters and self-generated code skills.
This tutorial surveys 3D vision-language models, covering geometric representations (point clouds, meshes, Gaussian splatting), CLIP-style cross-modal alignment, taxonomy of embedding-based vs. generative architectures, and emerging applications in recognition, scene understanding, 3D generation, and embodied AI, highlighting the distinction between similarity matching and language generation capabilities.
3D Gaussian splatting3D vision-language modelsCLIP
This article explains world models as AI systems that simulate physical world rules, contrasting them with LLMs' statistical pattern matching, detailing their capabilities in reasoning, prediction, and cross-domain generalization, while noting high training costs and current limitations in complex scenarios.
Zhejiang University's APRIL Lab publishes a comprehensive survey establishing a systematic framework for smart glasses as first-person intelligence platforms, defining eight hardware capability axes, seven fundamental capabilities, an L0-L5 maturity model, nine application scenarios, and a nine-dimensional deployment framework with evaluation protocols.
Artificial Analysis's AA-Robotics benchmark tests GPT, Claude, and Gemini on real robot arm tasks like pouring and stacking, finding they succeed but take ~30 minutes per task at high cost, with low grasp success rates, highlighting gaps versus specialized embodied AI models.
Amap had 16 papers accepted at NeurIPS 2026, covering multimodal large models, spatial intelligence, visual reasoning, and reinforcement learning, with highlights including a decomposed verifier for generative world modeling, gradient projection for continual multimodal learning, map-free transit route generation, and benchmarks for coding agents and travel planning.
3D ReconstructionEmbodied AIMultimodal Large Models
RoboFoundry freezes foundation models and evolves the surrounding system — context, skills, and execution interfaces — via a closed-loop of execution, reflection, repair, and promotion, achieving large gains on EmbodiedBench, RoboMemArena, LIBERO-PRO, and real-robot zero-shot transfer and online evolution.
USTC spinout Baize Tongjing develops Abductive World Model (AWM) that predicts then reasons about causes, learns from unlabeled video with 7x data efficiency, and deploys via 2-bit quantization engine BAIZ-RUN on edge chips, validated in three industrial robot scenarios within three months of founding.
Researchers from Renmin University and Beijing Academy of Artificial Intelligence introduce ROMA, a system that equips robots with multisensory active perception, enabling them to autonomously decide which object to interact with, how to interact, and which modalities to use, forming a reasoning-interaction-feedback loop that matches GPT-6 Astra performance on real-world tasks.
The EmbodiedJev open-source project integrates the Jev fast decision-making model into a MuJoCo-simulated Franka Panda robot arm, enabling browser-based control for pick-and-place, block stacking, and obstacle avoidance tasks with support for multiple LLMs and three decision-making modes.
Researchers from Griffith University and Tsinghua University introduce Spotter, a robot agent framework where an embodied model executes tasks while a vision-language model supervises in parallel, intervening only upon detected errors, achieving minute-level task completion with 70% less time than VLM-led baselines across RoboCasa, RoboTwin 2.0, and real-robot experiments.
Researchers from 15 institutions introduce TouchScale, a 500-hour human vision-tactile dataset collected with a single wearable system, demonstrating that scaling tactile data under consistent hardware and procedures yields cross-sensor generalization gains and lifts real-robot manipulation success from 22.5% to 57.5% across four contact-rich tasks without using human action labels.
Sergey Levine explains why robotics must prioritize decision-making over hardware, leverage prior knowledge instead of tabula rasa learning, and build diverse real-world data flywheels before scaling—highlighting that the field hasn't yet found its 'Transformer moment' for reliable generalization.
The article examines how embodied intelligence data competition is shifting from sheer volume to measurable capability gains, detailing Real2Sim2Real closed loops, world models, and commercial validation showing simulation success rates jumping from 9.7% to 79.8% and real-world task success improving over 50%.
RoboICL enables frozen GPT-6 Astra to control dual-arm robots by providing demonstrations and the robot's own execution history as embodied context, achieving 50.64 average progress score on 30 RoboDojo tasks—surpassing prior methods—with detailed analysis of memory design, real-robot transfer, and 929 executable trajectories for inspection.
Show-Harness introduces an Embodied Harness — a semantic action interface enabling vision-language models to control diverse robots through interpretable, composable actions like move, rotate, grasp, and release, achieving 89–100% zero-shot success across tasks, environments, and embodiments while unifying human teleoperation, GUI agents, and VLM control via the GUMI interface.
This article introduces articulated objects — multi-part 3D structures connected by joints like hinges and sliders — explaining their core components (links and joints), common joint types (revolute and prismatic), and why understanding their motion is crucial for robot manipulation, digital twins, and interactive VR/AR environments.
At the 2026 Yunqi Conference, Banma Smart unveiled AutoOmni 2.0-23B-A3B, a sparse MoE edge model that achieves near-cloud performance on vehicle-grade chips, addressing memory, latency, and concurrency constraints while enabling privacy-preserving, low-cost AI with certified safety frameworks for mass production.
Digua Robot leverages Volcano Engine's multimodal data lake — combining TOS object storage, Lance columnar format, and Ray distributed processing — to unify heterogeneous robot episode data, enable PB-scale ingestion, spatial quality checks via LAS operators, and achieve near-PFS training throughput with RobotLanceDataset.
The article reports on the "Triple T" Zhangjiang Top Meeting salon where six experts presented on recursive self-improvement (RSI) in AI, covering agent-environment-data-model co-evolution, self-evolving code models, discovery intelligence, embodied AI, AI for science, and a roundtable on achieving genuine RSI beyond local optimization.
AI Self-EvolutionAI for ScienceAgent-Environment Co-evolution
Baidu Baige optimizes DataLoader across sampling, indexing, decoding, and post-processing for embodied model training, reducing FastWAM single-sample video processing from 133ms to 63ms, stabilizing indexing at 0.61ms across 27,500 files, and boosting Qwen3-VL-30B throughput 1.68x via load-balanced sampling.
This daily observation reports AI applications entering mass deployment across manufacturing, healthcare, transportation, agriculture, finance, governance, and consumer sectors, highlighting humanoid robot factories producing 10,000+ units annually, a medical imaging AI published in Science, banking AI agents handling billions of tokens, and unified government AI platforms.
Banking AI AgentsEmbodied AIGovernment AI Platform
This daily observation report details 18 real-world AI application cases across eight sectors—manufacturing, healthcare, agriculture, finance, low-altitude economy, urban/rural governance, embodied terminals, and investment promotion—showing how AI has moved from conversational tools to production-grade deployment with measurable efficiency gains, cost reductions, and societal benefits in mid-September 2026.
AI industry applicationsEmbodied AIagricultural AI
Researchers test GPT-6 Astra as a general-purpose robot brain, achieving 95% success on block pick-and-place but only 10% on precise puzzle insertion, revealing that while large language models can plan robot actions, fine-grained physical control remains a challenge for pure AI-driven systems.
Tokyo-based Enactic_ai fully open-sources OpenArm, a 7-DOF humanoid robotic arm including 3D-printable CAD files, BOM, motor firmware, ROS2 drivers, web-based MuJoCo simulation, Isaac Sim integration, and a leader arm for zero-latency force-feedback teleoperation, dramatically lowering the cost barrier for embodied AI research.
The article reveals that synthetic data, once touted as AI's infinite fuel, now costs up to thousands of yuan per high-quality reasoning sample due to MCTS exploration, formal verification, and expert review, while hitting diminishing returns from model collapse, boundary limits, and compute-cost divergence, pushing the industry toward embodied interaction and formal environments.
This article analyzes OpenMind's OM1, a Go-based modular AI runtime and hardware abstraction layer for robots that uses a natural language data bus to coordinate multiple LLMs, enabling hardware-agnostic deployment across humanoid, quadruped, and simulated robots with a plugin architecture and JSON5 configuration.
Researchers from Shanghai Jiao Tong University and Shanghai AI Lab introduce REAL, an ECCV 2026 framework that enables embodied agents to actively explore unknown environments, disambiguate vague user instructions through dialogue, and execute mobile manipulation tasks via a unified MCP tool interface, achieving 78.3% real-world success on a dual-arm robot.
Stanford professor Jiajun Wu's ECCV 2026 talk explores how structured world representations enable robots to act in complex physical environments, covering compositional skill learning, neural kinematics for zero-shot object interaction, video diffusion models for generating free demonstrations, and new benchmarks for evaluating embodied reasoning.
Light Source Innovation, founded by ex-OpenAI RLHF expert Jiang Xu, releases LightNav-0 for zero-shot cross-morphology navigation and Light REACT for whole-body resilience control, applying LLM-style scalable pre-training, alignment, and deployment paradigms to embodied AI with sim-to-real synthetic data and preference-aligned RL.
Liangyuan Xinchang releases LightNav-0, a zero-shot navigation model deployable across robot morphologies, and Light REACT, a resilient whole-body control policy that recovers from falls and damage, applying the LLM pre-training/alignment/deployment paradigm to embodied AI.
PhyAgentOS v1.0.0 introduces an open-source harness that unifies heterogeneous robots, verifies task outcomes with evidence, enables bounded recovery from failures, and evolves skills through verified experience, achieving 98.6% success on LIBERO and measurable gains on CALVIN and RoboCasa benchmarks.
This survey reviews vision-language models for egocentric video, analyzing why standard VLMs fail on first-person footage and how hand-object interaction, spatiotemporal reasoning, graph-based representations, and semantic alignment bridge the gap toward embodied AI applications.
HiDream.ai's new embodied world model HiDream-O1-Embodied achieves first place on the RoboColiseum robustness benchmark with a 0.692 score, demonstrating superior language understanding, multi-view visual perception, and fault tolerance through a native multimodal architecture and a novel 'real base + generative augmentation' data paradigm, completing the company's unified world model matrix.
The article analyzes how Chinese robotics firm Xinghaitu uses Gemini, Doubao, and SAM3 to automate embodied data annotation pipelines, detailing the multi-step process from video segmentation to trajectory projection, and discusses the shift of human roles to quality control, the risk of instruction-trajectory mismatch poisoning VLA models, and four capabilities annotation companies must build.
Meta's HumanCLAW benchmark evaluates nine vision-language models on embodied action intelligence, separating high-level decisions from low-level control; the best model completes full interactions at just 16.8% success, revealing critical gaps in embodied self-awareness and closed-loop reasoning.
This article analyzes EmbodiSkill and ASPIRE research to derive three design principles for continuous evolution of embodied agents: fine-grained feedback attribution to distinguish skill defects from execution errors, skill libraries as core knowledge accumulation substrates, and multi-candidate exploration with competitive validation to improve robustness.
Researchers from Shanghai Jiao Tong University, NUS, Meituan, and others built UrbanGround, a real-scale interactive Hong Kong environment using 3D geographic data, to evaluate MLLM agents on local perception, navigation persistence, and dynamic adaptation, revealing models excel at local tasks but fail at long-range navigation and dynamic recovery.
This article summarizes four ECCV 2026 papers: a dynamic cross-layer injection framework for deep vision-language fusion, an event-augmented VLA model enabling robot operation in extreme darkness and blur, a predictive differentiable rendering method using 2D Gaussians for high-fidelity video prediction, and a data-scaling approach for high-resolution weather forecasting that demonstrates clear scaling laws.
The paper introduces TrAct, which replaces action‑conditioned world models with visual‑track conditioning, letting a policy output both robot actions and 2‑D visual trajectories that guide a future‑prediction model, and demonstrates substantial gains on the LIBERO‑INTEGRAL benchmark and real‑robot tests.
The article analyzes the severe shortage of real‑world robot‑body interaction data, quantifies the gap between existing high‑quality embodied datasets and the massive hours needed for general models, and explains how industry players and novel legged platforms are trying to bridge this gap.
GigaBrain-0.7, the latest embodied foundation model released by JiJia Vision, introduces the novel System‑3 paradigm that embeds a world model into the robot's decision loop, leverages a dual "data pyramid + algorithm pyramid" framework, and achieves top rankings across all RoboColiseum sub‑benchmarks, demonstrating unprecedented real‑world performance and scaling behavior.
The article analyzes how Gaode Momentum’s TuTu robot dog overcomes high deployment costs by using map‑free spatial understanding, multimodal perception, and a unified embodied AI brain to enable reliable autonomous navigation and task execution across five diverse real‑world scenarios.
The article analyzes how Lumos NexCore aims to transform robot skill production from a manual, project‑by‑project process into a cloud‑like, reusable infrastructure, detailing the platform's five‑stage pipeline, industry challenges, competitive shifts, and the broader impact on embodied AI deployment.
Zetta ζ introduces a three‑level closed‑loop system that lets robots monitor, recover, and update skills online, turning failure‑prone static agents into self‑evolving systems that achieve jump‑start improvements—from 15% to 95% success on simple tasks and over 20‑point gains on LIBERO‑Pro and RoboCasa benchmarks—without retraining the underlying policy model.
The article explains a three‑layer Skill architecture for AI agents, shows how real user feedback is turned into concrete rule updates across routing, instruction, and resource layers, and describes iterative refinement, compaction, and validation before releasing new Skill versions.
The RoboColiseum platform tackles the lack of reproducible embodied AI benchmarks by offering high‑fidelity simulation with 89.5% real‑world alignment, four‑dimensional evaluation across 78 tasks, rapid 30‑minute assessments, baseline model comparisons, and support for over 40 global research teams.
The article explains why traditional robot log formats struggle with self‑organizing behaviors, compares the mainstream standards MCAP, ROS Bag 2.0 and ULG, and evaluates four visualization tools—PlotJuggler, Roboto, robot‑log‑visualizer and WandB—showing how they support efficient recording, storage, and analysis of multimodal robot data.
The article analyzes why traditional model deployment fails for embodied intelligence, outlines three prevailing solution routes, and presents Lamp—a unified compiler framework that transforms complex robot pipelines from Python prototypes into edge‑ready, high‑performance products, achieving up to 22 Hz inference and offering five integrated capabilities.
Daimon‑TWM, the world’s first tactile‑grounded model, combines massive tactile data, perception‑to‑reasoning pipelines and fast‑feedback control to let robots predict and adapt to physical interactions, achieving dramatically higher success rates than vision‑only or prior tactile models, even under disturbances.
The paper proposes a five‑layer Data Pyramid for embodied manipulation, analyzing each layer—real‑robot, UMI‑style, ego‑centric, simulation, and generic vision‑language data—through quality, diversity, reusability and physical fidelity dimensions, and discusses how these layers can be combined for training embodied AI models.
The article examines Superdimensional Power’s end‑to‑end embodied AI pipeline—from massive first‑person human data collection and a three‑stage training process to high‑DOF humanoid robots and world‑model generation—highlighting technical challenges, hardware‑algorithm coupling, and efficiency metrics that determine whether a cloud‑brain can reliably empower diverse robots.
The HOST framework lets a robot acquire a brand‑new skill by watching a single 29‑second human demonstration video, achieving a 62% success rate, 500× faster learning, and 50× lower data usage while preserving existing abilities.
RoboHarness demonstrates that no single embodied model can handle all long‑horizon robot tasks; by dynamically selecting and bridging between VLA, RL, TAMP and other policies using Understanding, Memory, and Evolution skills, it achieves up to 95.2% success on challenging LIBERO benchmarks.
The article explores how Jisu Tech, a pioneer in data‑quality for embodied AI, builds a ‘refinery’ for spatial data by turning measurement expertise into a quality‑assessment model, a factor library, and a data‑flywheel that fuels autonomous‑driving and future intelligent‑robot applications.
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 analyzes Real2Sim's approach of constructing simulation environments directly from massive, millimeter‑accurate 3D scans, highlighting its zero‑error reconstruction, multi‑modal data richness, scalable scene generation, and how it surpasses video‑driven methods like SimFoundry for embodied AI training.
Muka Robotics' embodied world model LJM achieved second place in the WorldArena leaderboard with an EWMScore_P of 73.06, thanks to a dual‑expert architecture, MoT shared attention, Value‑Driven Temporal Conditioning, and full‑process training on 32 Alibaba Cloud Zhenwu 810E GPUs, which also delivered state‑of‑the‑art results on the LIBERO benchmark.
The TriWorldBench Challenge, launched by top Chinese universities, introduces a three‑camera evaluation suite for embodied world models that tests multi‑view consistency, task execution, and physical understanding across 500 synchronized episodes, providing a diagnostic TWB‑Score to guide future research.
The article presents a 30,000‑hour multimodal tactile‑visual dataset, unified representation models, and two downstream systems (VTLA and TWAM) that demonstrate how large‑scale touch data can dramatically improve robot manipulation success rates and establish tactile perception as a core scaling law for embodied intelligence.
The CD‑LAM method removes visual contamination from latent action models, cutting action error by over 30% and boosting image quality, while requiring just 3‑6k fine‑tuning steps (about 10% of baseline cost), as demonstrated on 2B and 14B scale world models.
The article analyzes the limitations of reactive VLA models for long‑term robotic tasks and presents τ0‑VLA, a hierarchical “slow‑thinking, fast‑execution” system that integrates world‑model‑guided test‑time computation, large‑scale real‑world data, and a unified 40‑dimensional action space, achieving significantly higher success rates on multi‑step tasks.
NeoteAI and Fudan University release a 30,000‑hour visual‑tactile dataset and three models—NeoForce, VTLA and TWAM—that demonstrate tactile scaling, predictive touch for VLA, and multimodal world‑model integration, achieving up to 99% task success and proving touch as a core building block for embodied intelligence.
The article details how Yingmu Technology’s cuNRTO paper, nominated for the RSS 2026 Outstanding Paper Award, moves nonlinear robust trajectory optimization onto GPUs, delivering up to 139.6× speedup while preserving 100% safety constraints, and situates this breakthrough within the company’s broader 3D‑to‑embodied‑AI research roadmap.
The paper introduces Harness VLA, a system that adds a Harness Layer to frozen Vision‑Language‑Action models, uses an Agentic Planner for task orchestration and failure recovery, and achieves 82.4% success on the challenging LIBERO‑Pro benchmark—far surpassing Pi_RLinf (50%), NVIDIA Cap‑X (18.2%) and Berkeley RATS (43.8%).
At this year’s WAIC, MechaMind showcased a suite of robots—humanoid, wheeled, and arm‑based—all driven by a shared embodied AI "eye‑brain‑hand" system, demonstrating how a single multimodal model can generalize across bodies, tasks, and environments while meeting industrial speed, precision and reliability demands.
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.
At WAIC 2026, JD unveiled its JoyAI model matrix, open‑sourced the largest human‑viewpoint dataset, and demonstrated an AI‑powered “AI Home,” illustrating a complete pipeline that moves AI from digital screens into real‑world perception, decision‑making, and action.
At this year’s WAIC, JD showcased a suite of embodied AI models—including JoyAI‑Image‑Edit, JoyAI‑Video‑Edit, JoyAI‑Voice, and JoyAI‑RA—demonstrating how AI is moving from screen‑based perception to real‑world sensing, decision‑making, and actuation, backed by massive first‑person data collection and a closed‑loop training pipeline.
The 2026 WAIC showcased over 200 robot firms, but the real competition now hinges on who can build a closed‑loop physical AI system that continuously captures, learns from, and deploys real‑world experience at scale, a challenge JD.com is tackling with its JoyAI ecosystem.
Visics demonstrated the world’s first cloud‑based embodied AI model at WAIC 2026, showing a single brain controlling multiple robotic hands that can be swapped in 30 seconds without retraining, achieving 99% grasp success across ten hand types using a VLOA architecture and massive video‑simulation data.
Astribot released the Lumo‑2 embodied model, showcasing over 20 real‑world household tasks—from collaborative box‑folding to fine‑grained coffee‑making—while introducing a latent world‑action architecture, three‑stage cross‑modal alignment, a 2.71× faster inference engine, and the modular Agent Philia system that together illustrate a full‑stack AI‑OS‑body approach poised to reshape home robotics.