Skip to content
#Model releases
Oct 28Wed
  1. 机器之心 · 微信公众号73

    Google DeepMind releases world model Genie 4, supporting real-time control and native audio

    Google DeepMind has released Genie 4, its fourth-generation world model, which can generate dynamic environments in real time; users can control characters or the camera with the keyboard and use natural language to change the weather and add objects and characters. Genie 4 is Google's first world model to natively generate real-time audio, and it also improves physics simulation such as aerodynamics and collisions as well as the naturalness of character movements.

  2. 雷科技56

    Google releases Genie 4 world model: supports real-time interaction, can generate robot training scenarios

    Google has released a research preview of its world model Genie 4, which natively supports multimodality, can change the environment and introduce new objects and characters through instructions during generation, and lets existing characters react. The model is integrated with Gemini, with Gemini Omni anchoring the simulated visual behavior logic to the real world. It is currently open only to invited testers and will later open first to Ultra subscribers.

  3. 新智元 · 微信公众号80

    Google releases world model Genie 4, with three Gemini models jointly driving robot actions

    Google DeepMind has released Genie 4, a new-generation general-purpose world model that natively generates sound and improvised dialogue for the first time, sustaining a single coherent world for about 2 minutes at roughly 24 frames per second of real-machine footage. The model uses a hybrid world-model architecture: Gemini 3.8 Flash handles real-time control, while Gemini 4 Argon breaks tasks into 22-step action instructions, driving an Apollo robot to play checkers and, in simulation, controlling a UR5 robotic arm to perform pipetting operations.

    Why it matters: Google has released the Genie 4 world model, with three Gemini models jointly driving robot action decomposition; readers can learn how world models are used for robot training.

  4. CVer · 微信公众号78

    Google releases world model Genie 4, supporting real-time generation and collaborative reasoning with Gemini

    Google DeepMind has released Genie 4, a new-generation world model that supports real-time generation of explorable virtual worlds, simulates physics and character movements, and natively generates ambient sound effects and dialogue. Genie 4 connects to Gemini 3.8 Flash as a real-time guidance agent, with Gemini planning tasks, observing feedback, and adjusting strategies. In the demo, it simulated robotic arm experimental operations and the Apptronik humanoid robot Apollo playing checkers in the generated environment.

    Why it matters: Google released the Genie 4 world model, demonstrating its ability to collaborate with Gemini to plan and simulate robots executing tasks in virtual environments.

  5. 智东西62

    iFlytek Releases SparkAudio 1.0 Speech Foundation Model

    During its 2026 Global 1024 Developer Festival, iFlytek released SparkAudio 1.0, a speech foundation model trained entirely on domestic compute. It supports text and speech dual-modal input with text output, and can perform speech transcription, multilingual translation, multi-dialect recognition, environmental sound recognition, speaker separation, sentiment analysis, and complex audio question answering.

  6. 芯东西73

    Google releases world model Genie 4, capable of simulating aerodynamics and collisions with native audio generation

    Google released its next-generation world model Genie 4 on October 28, capable of simulating physical processes such as aerodynamics and collisions, and natively generating music, ambient sound, and interactive dialogue. The model combines capabilities from the Gemini series, with Gemini 3.8 Flash breaking down multi-step tasks in a loop, and Google also demonstrated it combined with Gemini Robotics to simulate an Apptronik Apollo humanoid robot playing checkers.

  8. The Humanoid Hub75

    Google DeepMind releases Genie 4 world model, capable of generating robot training environments

    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.

  10. Shanghai AI Laboratory - Embodied Intelligence60

    Shanghai AI Laboratory and partners jointly launch materials multimodal large model "Caihua·MatVerse"

    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.

Oct 27Tue
  2. Kinetix AI · YouTube67

    Kinetix AI releases KAI World Model, supporting locomotion and manipulation tasks

    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.

  3. 雷科技70

    AgiBot releases embodied Agent framework Genie Master, validating Physical RSI in a real convenience store scenario

    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.

  4. arXiv cs.RO77

    AgiBot Team Releases GE-Act 2.0 World-Action Model Technical Report

    The AgiBot Research Team released a technical report on GE-Act 2.0, a world-action model whose generative and action components are all trained from scratch on manipulation data. It consists of a control-oriented autoencoder (CoAE), a single-step visual planner (SVP), and an inverse dynamics model (IDM), and is jointly trained with KASO.

    Why it matters: The original text presents the component design and data-scaling curves for pretraining a world-action model from scratch, allowing observation of changes in zero-shot cross-embodiment transfer.

  5. arXiv cs.RO63

    RoboJev-27B: An Open-Source Vision-Language Model for Judging Robot Task Success

    RoboJev-27B is an open-source vision-language model that takes a task instruction and 10 frames of a manipulation attempt as input and predicts whether the robot completed the task. It achieves 95.2% balanced accuracy and 0.993 AUROC on 268 held-out episodes of the same robot model (SO-101), and an AUROC of 0.890 on 459 episodes from unseen labs or robots. Median inference time is 0.26 seconds per episode on a single H100, and the model, inference code, and evaluation results for 20,628 predictions are released together.

Oct 26Mon
  1. 机器之心 · 微信公众号75

    COCO Matrix releases robot model CO01, using experience prompts to let robots learn while doing

    COCO Matrix released the robot model CO01, which treats demonstrations, failed attempts, and real-time guidance as experience prompts in context, allowing adaptation to new tasks without updating weights. Under the same pretrained model, a roughly 1-minute demonstration achieved an average success rate of 52% on unseen tasks; fine-tuning to a similar level required about 180 demonstrations and about 9 hours. In real-robot evaluations with at most five attempts per task, cumulative success rate rose from 33% on the first attempt to 84% when retaining its own history, versus 56% without retaining history.

  2. Robotics & Automation News78

    Dream Labs raises $90 million seed round and launches Dream API and Dream-1 model

    Dream Labs announced it has completed a $90 million seed round led by B Capital, with participation from Kindred Ventures and others. The company also launched Dream API, which can convert robot camera footage and natural-language instructions directly into motor commands, and said its flagship model Dream-1 is trained on videos of everyday human tasks and can zero-shot control any robot with a gripper.

    Why it matters: Dream-1's cross-hardware zero-shot control and the open approach of Dream API provide a concrete reference for judging the transferability of general-purpose robot models.

  3. arXiv cs.RO72

    RynnUnify Releases 16B Unified Embodied Model Combining Visual Planning and Robot Control

    RynnUnify is a 16B unified embodied model that combines embodied understanding with long-horizon visual planning through interleaved video-language reasoning, covering spatial reasoning, physical prediction, and grounding. The authors also developed an agentic extension based on reinforcement learning to selectively use video prediction to reason about unobserved states and interaction outcomes, and a world-action extension that translates understanding and prediction capabilities into robot control.

Oct 24Sat
  1. 机器之心 · 微信公众号72

    Limpid Intelligence releases PrismWAM and RoboDoctor, topping RoboTwin 2.0 and RoboDojo leaderboards respectively

    Limpid Intelligence founder Li Yiming's team at Tsinghua University released the world action model PrismWAM and the execution diagnosis framework RoboDoctor. PrismWAM decouples semantic context from physical state and, without additional action-data pretraining, achieved an average success rate of 83.67% on RoboTwin 2.0 and 75.5% in the Clean-to-Random setting, ranking first on the official leaderboard.

  2. HuggingFace Daily Papers71

    RoboJev-27B: An Open-Source Vision-Language Model for Robot Manipulation Success Detection

    RoboJev-27B is an open-weight vision-language model that takes a task instruction and 10 frames of a manipulation attempt to judge whether the robot completed the task. It achieves 95.2% balanced accuracy and 0.993 AUROC on 268 held-out episodes of the same robot model (SO-101), and an AUROC of 0.890 on 459 episodes from unseen labs and robots. Median inference time is 0.26 seconds per episode on a single H100, and the model, inference code, and evaluation results for 29 model/protocol entries are released together.

Oct 23Fri
  1. 新智元 · 微信公众号58

    ZhiZhi Innovation Research Institute publishes IQuest-Q1 technical report: training details of a 320B open-source base model on one million tasks

    ZhiZhi Innovation Research Institute has published a complete technical report on IQuest-Q1, which has 320B total parameters and 15B activated parameters, disclosing results on 30 benchmarks and training details. The team built more than 1 million executable, containerized task instances from code repositories and existing agent trajectories, and set up a three-layer mechanism of environment preparation, runtime checks, and trajectory auditing to prevent reward hacking.

  2. 钛媒体62

    Capital rushes into AI 3D: VAST, Meshy, and Yingmou absorb over RMB 8 billion in half a year

    FoST Future Narrative published an analysis of the financing and commercialization landscape in the AI 3D sector: as of the end of September, three companies founded by people born in the 1990s—VAST, Meshy, and Yingmou—had raised a combined total of over RMB 8 billion within half a year. VAST raised over RMB 5 billion in half a year and reached a valuation of USD 2 billion; Meshy completed a USD 400 million financing round in July 2026 at a post-money valuation of USD 1.5 billion, and its ARR surpassed USD 100 million at the end of September.

  3. HuggingFace Daily Papers62

    Ψ-0.6: Adaptive Streams of Thought in a Unified Vision-Language World Model

    The research team proposes Ψ-0.6, a unified vision-language world model that adaptively combines perception, language, geometry, and simulation within a shared reasoning process called streams of thought. These streams alternately generate and interpret images, text, and explicit representations of geometry and motion through a shared token interface. The model surpasses leading open-weight models zero-shot on multi-view spatial reasoning, with further gains from reasoning through geometric intermediate representations.

  4. HuggingFace Daily Papers75

    RynnUnify releases 16B unified embodied model combining visual planning and robot control

    RynnUnify is a 16B unified embodied model that combines embodied understanding with long-horizon visual planning through interleaved video-language reasoning, covering spatial reasoning, physical prediction, and grounding. The authors also built an agentic extension using reinforcement learning and introduced a world-action extension that translates understanding and prediction capabilities into robot control.

  5. 36Kr Embodied Intelligence55

    iFlytek releases device-cloud integrated Spark X2.6 and speech foundation model SparkAudio 1.0

    On October 22, at the main forum of the 2026 iFlytek Global 1024 Developer Festival, the device-cloud integrated iFlytek Spark X2.6 was released, with capabilities on both the cloud and device sides (Spark X2.6-4B) improved in sync, balancing task performance, security, and cost through precise division of labor and on-demand collaboration.

  6. 机器人全球资讯 · 微信公众号62

    XPeng IRON humanoid robot upgraded: re-reasons every 0.2 seconds, can predict motion changes within 6 seconds

    XPeng announced that its second-generation VLA large model, after an upgrade, has been deployed on-device to the IRON humanoid robot, giving it temporal understanding and future behavior prediction capabilities. IRON re-reasons every 0.2 seconds based on the scene before it and predicts motion changes within 6 seconds, allowing it to respond to last-minute changes during handshakes, fist bumps, and high-fives. XPeng plans to achieve mass production of humanoid robots by the end of 2026, begin deliveries in the first half of 2027, and first deploy them in commercial scenarios such as stores and shopping malls.

  7. Robotics & Automation News72

    Google DeepMind uses Gemini 4 Argon to teach Apptronik's Apollo humanoid new skills from video

    Google DeepMind demonstrated a robot training technique alongside the release of Gemini 4 Argon: after watching raw video of a person playing cornhole, the model analyzes the environment, maps spatial coordinates, and computes full-body joint motion, automatically generating simulation code, then transfers the learned behavior to Apptronik's Apollo humanoid robot, which hit seven consecutive shots in real-machine testing.

  8. 甲子光年 · 微信公众号66

    RockAI Matrix open-sources EdgeRock-4B: KV Cache does not grow with context, single-card throughput reaches 6.42x the original model

    RockAI Matrix, under RockAI Technology, together with universities including the Hong Kong University of Science and Technology and Carnegie Mellon University, released and open-sourced EdgeRock-4B. The model is converted from Qwen3.5-4B, and its core is the Memory Rock Anchor (RMA) mechanism proposed by the team, which keeps the KV Cache from growing with context.

Oct 22Thu
  1. 高飞的电子替身 · 微信公众号72

    Jensen Huang GTC Berlin keynote: Vera Rubin mass production moved up, OpenShell security platform adopted by over 100 organizations

    On October 21, 2026, NVIDIA held its first Europe-hosted GTC conference in Berlin, where Jensen Huang delivered a keynote. Ahead of the keynote, NVIDIA announced that all seven chips in the Vera Rubin GPU family had completed tape-out, with mass production moved up to the third quarter of 2026; OpenShell, the open-source agent security platform released at the end of September, has already been announced as adopted by over 100 organizations; and the RTX Spark series of PCs was confirmed to launch this fall.

  2. 智东西66

    iFlytek releases end-cloud integrated large model Spark X2.6 and Robot Super Brain Platform V3.0

    iFlytek has released the end-cloud integrated large model iFlytek Spark X2.6, comprising the cloud-side 293-A30B model and the already open-sourced device-side X2.6-4B model. The cloud model supports a one-million-token context, with an input price of 0.9 yuan per million tokens and an output price of 3.6 yuan per million tokens, lower than DeepSeek-V4.1-Flash-1017.

  3. The Robot Report78

    Dream Labs raises $90 million and launches Dream-1 model and Dream API

    Dream Labs announced this week that it has raised a $90 million seed round and launched the Dream API, which converts robot camera feeds and natural-language instructions directly into motor commands. Its flagship model, Dream-1, is trained on videos of everyday human tasks; the company says it can control new gripper-equipped robots zero-shot without additional data. It currently supports tasks such as tabletop pick-and-place and clothes folding, with wheeled mobile robots and dexterous hands to be supported later.

  4. 钛媒体57

    Chery releases CheryMind, a vertical foundation model for physical AI

    At the 2026 Chery Global Innovation Conference, Chery released CheryMind, a vertical foundation model for physical AI. It serves as the technology foundation of the Yaoguang 2035 strategy, covering automotive R&D and manufacturing and extending to businesses including intelligent driving, intelligent cabins, robotics, eVTOL, and agricultural machinery.

  5. NVIDIA75

    NVIDIA Cosmos Expands with New Tools for Training, Evaluating and Deploying World Models

    NVIDIA has expanded the Cosmos platform with new training tools, simulation checkpoints and Cosmos Data Factory blueprints for training, evaluating and deploying world models.

    Why it matters: The original lists Cosmos's new training tools, simulation checkpoints and data blueprints, and gives specific data on users such as Porsche and ANYbotics, which can be used to assess changes in the world model development workflow.

  6. 仙工智能58

    SEER Robotics robotic arm autonomously completes long-horizon coffee-making task

    SEER Robotics demonstrated its robotic arm completing the full coffee-making process without human intervention, from picking up the portafilter, receiving grounds, distributing and tamping the grounds, locking the portafilter, extraction, to cleaning. This is the longest multi-step robotic coffee task the company has demonstrated to date.

  7. Ken Goldberg62

    Researcher reverse-engineers Gemini 4 Argon skill library and open-sources robotics harness

    Ken Goldberg shared the work of his postdoc Shuangyu0_0Xie: she found that Gemini 4 Argon performed well on spatial robotics queries, so she used Claude to reverse-engineer its skill library and implemented the robotics harness described in the Gemini 4 Argon report. The harness enabled Gemini to complete a zero-shot bimanual handover, and it is now open source.

  8. Pieter Abbeel64

    SuperMimic cuts GPU hours for humanoid full-body control training by over 1,000x

    SuperMimic trains a humanoid full-body control policy on 129,514 motions, surpassing the previous best result of 128 GPUs over 7 days with 1 GPU in 3 hours, reducing GPU hours by over 1,000x. Pieter Abbeel shared it and called it a game changer, noting that the key lies in using a world model faster than conventional simulation and in collecting far more signal than policy gradients through backpropagation through time. The project page is https://supermimic.github.io, and the demo is at https://supermimic.github.io/demo/.

  9. 未来机器人网 · 微信公众号61

    ACE Robotics releases embodied foundation model ACE-Brain-0.6, surpassing GPT-6 Astra on six of eight Grounding benchmarks

    ACE Robotics released ACE-Brain-0.6, a vision-spatial-anchoring-enhanced embodied foundation model focused on Grounding capability. It surpasses GPT-6 Astra on six of eight Grounding benchmarks, with an average score of 74.53 across seven public benchmarks (Astra: 73.68), and scores 94.13 on its self-built real-robot evaluation set RobotGrounding (2,591 questions, 14 task types; Astra: 91.10).

Oct 21Wed
  1. 机器人前瞻 · 微信公众号82

    NVIDIA open-sources robot data engine Flora, labeling cost about 1/125 that of GPT-6 Astra

    NVIDIA released and open-sourced the data engine Flora, which converts raw robot video into training data such as timestamped action descriptions, process rewards, and outcome rewards; it is already used to train the GR00T series models. Flora consists of an action description model, a temporal segmentation model, and a reward model, fine-tuned from the Qwen3.5 series of vision-language models, and can label 200,000 hours of data per day on 512 GB300 GPUs.

    Why it matters: NVIDIA's open-source data engine cuts robot video labeling costs to about 1/125 of a general-purpose large model and provides several benchmark comparisons, which can be used to assess the substitution space in the data labeling segment.

  2. 机器人前瞻 · 微信公众号72

    NVIDIA Open-Sources Cosmos3-H-Dreams, a Real-Time Generative Surgical World Model That Runs on a Single GPU

    NVIDIA has released and open-sourced Cosmos3-H-Dreams, a real-time generative surgical world model that can generate surgical imagery in real time based on robot operation commands, simulating instrument movement, tissue traction, and deformation, with support for laparoscopic cholecystectomy scenarios.

  3. NVIDIA40

    NVIDIA and UW Medicine release open protein design model RFD4-Proteina; Siemens Healthineers advances adaptive radiotherapy

    NVIDIA and the UW Medicine Institute for Protein Design launched the open model RFD4-Proteina, combining programmable protein design with search capabilities. In four benchmark tests reported in a research draft—protein binding, small-molecule binding, enzyme scaffolds, and DNA binding—computational design success rates and generation throughput were both higher than the evaluated baselines.