The Era AI Lab team proposes E0.5, a geometry-native world-action model built directly on 3D foundation models that treats 3D geometry as the native representation for world modeling and action prediction, and explicitly predicts future 3D states and actions.
The study proposes the Just Action Model (JAM) framework, which breaks down the differences between VLA and WAM into three separable choices: backbone initialization, the future-observation objective, and whether that loss mechanism is retained at inference. It uses a pretrained visual backbone plus an action expert, trained with an action loss and a foresight loss.
ThinkWorld is a VLA learning framework that represents each action thought as a Chain-of-Dynamics (CoD), a sequence of discrete motion tokens that captures the intermediate dynamics from the current world to a future state and serves as a common action representation across embodiments.
LimX Dynamics and ZINOVA mounted the torso module of the TRON 2 dual-arm embodied robot on the end of a large six-axis industrial robotic arm, completing formwork frame assembly and rebar tying tasks at a heavy-industry construction site. The article notes that bipedal robots have a single-charge battery life of only one to two hours, with most of the power used to maintain balance, and that their small contact area makes it difficult to withstand the recoil of tools such as nail guns. In this approach, the industrial arm bears the load and absorbs impact, while the TRON 2 dual arms use vision to recognize an off-the-shelf industrial nail gun and plan gripping and trigger actions; force sensors perceive contact vibration and resistance, and two robots can work together to align lumber and fire.
Researchers propose SO-101 Bench, a language-conditioned spatial manipulation simulation benchmark built on an Isaac Lab digital twin of a real SO-101 setup, comprising four task types, 138 objects, 916 evaluation episodes, and 3,629 simulated demonstrations.
AGIBOT's GO-3 Nano Preview topped the latest RoboDojo embodied intelligence benchmark leaderboard with an average score of 61.90, with another result of 56.42%. The leaderboard was disclosed on October 27; specific evaluation conditions and task details were not elaborated in the original text.
Google DeepMind has released its next-generation world model Genie 4, extending the length of a single interactive simulation from about one minute to two minutes and supporting native audio generation for the first time.
A simulated Franka Panda learns in PyBullet to pick up blocks and place them on a target pad using PPO from Stable-Baselines3; the project aims to understand PyBullet and the full reinforcement-learning robot training pipeline. Future plans include adding text or voice prompts so users can choose which block to pick up and where to place it; the code has been released on GitHub.
Google DeepMind has released Genie 4, a real-time photorealistic world model that can be controlled with natural language, adding native audio and improved physics such as collisions and aerodynamics. The model can generate infinitely variable, controllable training environments for robots and agents, and has been paired with Gemini 4 Argon and Gemini Robotics to have Apptronik's Apollo humanoid robot complete a checkers task step by step in simulation.
This is Genie 4, our most realistic and interactive world model yet. 🌐 Explore dynamic environments in real time and observe how scenes respond, emit sound, and how the people within them move, react, and speak.
Suzhou National Laboratory, Shanghai AI Laboratory, Shanghai Jiao Tong University and others jointly launched the "Caihua·MatVerse" materials multimodal large model and agent application system, which uses an "independent encoders plus unified decoder" architecture to align XRD spectra, electron microscope images and molecular structures into the same semantic space.
Why it matters: The materials present a multimodal unified representation architecture and 19-category disciplinary evaluation benchmarks, offering insight into how materials large models can move from retrieval to mechanistic reasoning.
Toyota is commercializing two decades of robotics R&D, planning to deploy 400,000 robots backed by $6.4 billion a year in investment for factory automation, with a dedicated team coordinated by its chief production officer. In April this year Toyota showcased Cue7, an AI-equipped humanoid robot; in July Walden Robotics, spun off from Toyota's AI robotics business, completed a $300 million funding round at a post-money valuation of $1.1 billion, and its general-purpose robots are already deployed at Toyota's North American factories.
According to the Financial Times, Toyota plans to deploy 400,000 robots, backed by $6.4 billion a year in factory automation investment, with a dedicated robotics team coordinated by its chief production officer. Walden Robotics, spun off from TRI, has completed a $300 million seed round at a post-money valuation of $1.1 billion; its wheeled-base, dual-arm robot has been performing production work at Toyota's North American factories since February.
In an interview with Jiazi Guangnian, Force Infinite chief scientist Chen Jiayu judged that general large models such as GPT-6 Astra will completely eat up the task planning (Planner) portion, so the team will no longer continue investing in some high-level memory mechanism research and will instead focus on execution capabilities in the physical world.
On October 26, 2026, Robot Industry Application held a closed-door embodied AI data workshop in Suzhou Industrial Park, with more than 20 frontline decision-makers from models, data, real-world deployment, and chip infrastructure participating. Data vendors believe demand for embodied data is far from saturated, but model vendors argued that startups cannot scale data indefinitely; those on the deployment side emphasized cost reduction, efficiency gains, and the logic of calculating returns, and discussed mass production and standardization of tactile electronic skin.
MoSense, a company providing whole-body multimodal fusion tactile solutions, recently completed a hundred-million-yuan Angel+ funding round, led by CDH Investments and Oasis Capital, with participation from LimX Dynamics, Loongson Venture Capital, CAS Star, Huashan Capital, Pudong Venture Capital, and a leading domestic embodied AI manufacturer.
USV announced that it co-led Machenta's $40 million seed round with Creandum, as the company also emerged from stealth. Machenta says its robotics foundation model is trained entirely in simulation, with no real-world data, and can learn a task from a single demonstration without additional training, adapting when the environment changes.
Why it matters: USV, as co-lead investor, explains why Machenta's one-demonstration learning is worth betting on, giving readers insight into the technical claims behind the funding.
At the 2026 China Marine Economy Expo, Sightbit Robotics showcased a dual-robot collaborative welding workstation for shipbuilding, a humanoid welding robot prototype, an intelligent residual-material scanning system, a steel plate pitting 3D inspection system, and the Kunwu industrial simulation platform. The SpeedWeld intelligent welding control system supports reverse modeling with 3D structured-light cameras to generate welding trajectories, and can reduce changeover and commissioning cycles by up to 90%. The latest SpeedWeld 2.9 version supports dual-robot collaborative, asynchronous welding, and independent welding modes.
Manycore Tech, Meshy, and Hyper3D discussed why Chinese AI 3D companies are leading at the ACC 2026 AI Creator Carnival. Manycore has launched the spatial generation model SpatialGen and is exploring the 3D generation model Aholo Lux3D, whose materials support up to nine channels; Hyper3D launched Agentic Mode, moving from Production Ready toward Agent Ready. The three companies believe that domestic high-quality 3D data, the production pipeline ecosystem, and talent accumulation are the main advantages.
Harbin Institute of Technology (Shenzhen), the National University of Singapore, and other institutions propose SkeLAM, built on Cosmos-Predict2.5-2B and trained on about 10 hours of robot interaction data, using robot skeletons to supervise latent actions in video.
Researchers from UIUC, Google, Stanford, UC Berkeley and CMU jointly published the review "Proactive AI for Superintelligent Agents," systematically mapping the research landscape of proactive AI and proposing a 6A framework composed of six capabilities: Awareness, Anticipation, Agenda, Arbitration, Action and Adaptation.
Ropedia released six studies at the first Ropedia Super Week, held October 19 to 23, centered on how first-person human video can be turned into training and execution information usable by robots.
Why it matters: The six studies collectively address how first-person human data can be converted into training and execution information usable by robots, and present cross-embodiment and real-robot comparison results.
FastHarness is an experience-driven robot execution framework that uses π0.5 to generate actions and GPT-6 Astra to review and correct them, adjusting the review interval according to task stage and past review results. Across four simulated manipulation tasks in RoboDojo, it achieved an average success rate of 97.19%, higher than the 80% reported for π0.5 + GPT-6 Astra in the GPT-as-Policy report.
The Institute for AI Industry Research (AIR) at Tsinghua University, together with Tsinghua's Department of Electronic Engineering and the University of Science and Technology of China, has proposed OxyGen, which uses unified KV Cache management to let an embodied model's action and language reasoning share prefix computation. The paper has been accepted to NeurIPS 2026.
Why it matters: The paper presents a unified KV Cache management scheme and measured data across multiple platforms, offering a reference for on-device embodied inference deployment.
RoboParty has released the full-stack open-source humanoid robot RP1, opening the entire chain to developers, from mechanical drawings and electrical systems to its self-developed ROMOMO joints, low-level control, simulation training, and deployment. RP1 is 135 cm tall, with 24 active degrees of freedom in the open-source version and 30 in the Pro version; peak torque ranges from 25 to 160 N·m. It uses 48V power and CAN-FD communication, and its accompanying software provides a 500 Hz control frequency and a 50 Hz policy update frequency.
On October 26, Zhongke Huiling (CASBOT) announced the completion of an A++ funding round worth hundreds of millions of yuan, jointly funded by six institutions: Zhilai Investment, Yonghua Investment, Maotian Investment, Yuanhe Chenkun, Guoke Kezhi, and Luning Investment. Founder Zhang Zhengtao plans to allocate 60% of the funds to embodied brain R&D and 30% to product mass-production and delivery.
Why it matters: Beyond the funding, the article provides measured cycle times and payback periods for shoe-making workstations as well as progress on the mine demonstration, which can be used to assess the economic model of vertical embodied-intelligence deployment.
MoSense Technology announced on October 26 the completion of a 100-million-yuan angel funding round, led by CDH Investments and Oasis Capital, with participation from LimX Dynamics, Loongson Venture Capital, CAS Star, Huashan Capital, Pudong Venture Capital, and a leading domestic embodied AI manufacturer, following earlier investments from Sequoia China, Gaorong Ventures, and AgiBot.
A Tsinghua University research team proposes SimpleTouch, which adds a tactile expert to the pretrained VLA π0.5 and uses single-stage training with only 50 demonstrations per task, without additional tactile policy pretraining or a separate alignment stage.
An ECCV'26 paper proposes MV2GF, a multi-view pedestrian detection method that uses the multi-view features and 3D point maps of the pretrained visual geometry foundation model DA3 to aggregate BEV features instead of perspective projection. Under the setting where training and test camera configurations differ, MV2GF achieves MODA on GMVD-D, MVPerception, and Wildtrack that is 4.6, 4.7, and 2.2 points higher than the previous best, respectively.
Synrise announced the completion of a Pre-A funding round exclusively invested by Yunqi Capital, with proceeds going toward the SynchroMind algorithm foundation, R&D for next-generation products, and global market expansion. Within one year of founding, the company's first product, the Synrise S5 bionic intelligent hydraulic knee, has entered mass production, completed validation deployments in Europe, and received orders exceeding ten million yuan from domestic and overseas markets. A next-generation AI active-power knee is now in R&D, aiming to move from compliant motion to active power output.
Why it matters: An approach to human augmentation starting from intelligent prosthetics, showing how an algorithm foundation can be reused across products and accumulate real gait data.
EmbodiedLab is a benchmark that requires robots to "acquire knowledge through action," covering 30 quantitative science experiments across 5 disciplines on 14 apparatus platforms. The parameters of each instance are resampled, so results cannot be known in advance but can be scored precisely.
Kinetix AI has released the KAI World Model, which is based on parallel video generation and 4D spatial modeling and can simultaneously generate virtual worlds and reconstruct spatial scenes. The model was pre-trained on millions of videos and supports pure locomotion, pure manipulation, and coordinated locomotion-manipulation tasks; it claims generalization across tasks and across robot embodiments. Live demonstrations include a robotic arm autonomously opening a beverage can and a humanoid robot playing real-time table tennis rallies.
Why it matters: The original shows live demonstrations of the world model on two types of tasks—a robotic arm opening a can and a humanoid playing table tennis—which can be used to understand its approach to generalization across tasks and embodiments.
AgiBot released the embodied Agent framework Genie Master on October 27, enabling robots to autonomously identify the causes of failures after deployment, adjust strategies, and accumulate experience. The framework uses the Embodied Execution Protocol (EEP) to unify four types of interfaces: capability, guidance, feedback, and continuity, and through the two routes of Harness RSI and Policy RSI, it retains experience in prompts and memory, and in action policy parameters, respectively.
The latest leaderboard from embodied intelligence benchmark RoboDojo shows AGIBOT's GO-3 Nano Preview taking first place with an average score of 61.90 points and an average success rate of 56.42%, surpassing GPT-6 Astra. In the three sub-dimensions of fine manipulation, long-horizon execution, and open semantic understanding, GO-3 Nano Preview scored 78.65, 68.21, and 61.71 points respectively, ranking first in all three.
The authors propose the world action model Video Prediction Policy 2 (VPP2), which uses event-level pretraining on large-scale manipulation video, distills the video model into a single-step visual planner, and connects an action module via a MoT architecture to learn implicit inverse dynamics.
The authors propose the IG-VLA framework, which lets a vision-language-action model imagine task-relevant future scene evolution in visual representation space and internalize the reasoning results into a compact Scene Gist Token via Scene Gist Memory, thereby skipping explicit future imagination at inference time.
The authors propose JRDB-AVR, an active visual reasoning benchmark built on real JRDB robotics data: after receiving a visual reasoning question, an embodied agent must request limited observations by timestamp and viewing angle, and evaluation examines both the final answer and the visual evidence supporting it.
FlatVPR proposes a plug-and-play geometric linear residual adapter that applies a residual transformation to features from foundation models such as DINOv2-ViT-S/14 and uses a Pullback Flatness Loss to reduce manifold curvature, so that descriptors between adjacent anchors can be reconstructed by linear interpolation. On the NCLT dataset, the adapter still delivers clear performance gains even when anchor spacing reaches 100 m and severe seasonal changes are present. Map construction is incorporated into an EM framework, split into a continuous M-step for manifold adaptation and an E-step for the anchor selection criterion.
An ICLR 2026 paper re-examines the question of whether differentiable simulators can give better policy gradients, arguing that bias from discontinuous dynamics is not the main obstacle. The authors propose DDCG, a lightweight test that switches estimators in nonsmooth regions; it requires only one hyperparameter and remains reliable with small samples. They also propose IVW-H, a per-step inverse-variance implementation that stabilizes variance in differentiable robot control tasks without explicit discontinuity detection. The paper concludes that estimator switching can improve robustness in controlled experiments, while in real deployment variance control is often more critical.
The authors propose ActECC, which addresses the difficulty of determining the valid range of action chunks and scheduling replanning requests in edge-cloud collaborative VLA inference through kinematics-guided usable range assessment and delay-aware replanning scheduling. Experiments show it reduces per-action latency by up to 27.4%, cuts cloud calls by up to 45.1%, and lowers average execution error.