Dyna Robotics released its robot foundation model Dyna-2 in August 2026. It uses a WAM (world-action model) architecture, replacing the VLA approach of Dyna-1, and was pretrained on 1 million hours of first-person human video with zero robot operation data.
On the first day of Alibaba's open-source release of the 2.4-trillion-parameter MoE model Qwen3.8-2.4T-A95B, the Zhongzhi FlagOS community completed multi-chip adaptation, precision alignment, and deployment validation across 9 AI chips, providing BF16, FP8, INT8, and other multi-precision versions.
Why it matters: The article presents a zero-code-change adaptation path and a quantization compression scheme, which readers can use to assess how existing domestic compute can support ultra-large models.
Daimon Robotics released its tactile-anchored world model Daimon-TWM on August 10, and completed a strategic financing round of several hundred million yuan the next day, its second hundred-million-yuan-level financing within two months.
Higher task success rates and faster execution. 🚀 Our physical AI system completed a multi-stage unboxing task involving tool use and high-precision two-handed manipulation. Powered by #STEAM, it learns frame-level task progress from real-world experience of mixed quality.
In a StridingAI demo, the STEAM model from Professor Chao Yu's team at Tsinghua University drives a robot to autonomously complete a multi-stage unboxing task, achieving millimeter-level bimanual coordination through tool use. STEAM models task progress and action quality frame by frame, extracting learning signals from mixed-quality data including expert demonstrations, model rollouts, and human takeovers to improve task success rates and execution speed.
Dyna Robotics introduces Dyna-2, a world-action model pretrained on one million hours of human video. The authors say they found several new scaling laws at this scale: world-action models across four orders of magnitude on human data (from 1,000 hours to 1.
Why it matters: The original text provides data scales spanning four orders of magnitude and three scaling law conclusions, which can help readers understand how human video pretraining transfers to robot data.
X Square Robot has released and open-sourced the HOST framework, in which a robot can reproduce a new skill after watching an average 29-second human demonstration video, achieving a 62% success rate with only 1/50 the data and 1/500 the learning time of the Pi-0.5 fine-tuning approach. HOST does not imitate human actions frame by frame; instead, it uses visual outcomes as an intermediate language, aligning video with robot execution by task progress rather than time. It adopts a dual-expert MoT architecture, with learning occurring at the inference stage while weights stay frozen, and it retains old skills better than the fine-tuning approach; under lighting and scene changes, the success rate remains above 50%. The paper and code have been open-sourced.
RoboParty Lab, Zhejiang University, Tsinghua University Institute for AI Industry Research (AIR), and InSpatio jointly released and open-sourced the world model control method INTACT (INtent-To-ACTion).
Why it matters: The original article provides latency and success-rate comparisons with traditional CEM search, which readers can use to judge the benefit boundaries of search-free world model control.
Google has released Gemini Robotics 2.0; robots driven by the Gemini AI model can complete more complex tasks, continuously analyze changing environments, and collaborate with other robots. This version includes three new sub-models, among which the upgraded embodied reasoning model Gemini Robotics ER 2 is described by DeepMind as a significant improvement over the previous 1.6 version and has been integrated with the Gemini Live API; another sub-model is available to developers starting that day.
Google DeepMind has released Gemini Robotics 2, and robotics company Apptronik shared its congratulations. The quoted content describes it as a new generation of physical AI that brings whole-body intelligence to humanoid robots, with advanced dexterous manipulation and multi-robot collaboration capabilities. Apptronik said that seeing Apollo 2 use advanced reasoning to complete whole-body tasks and achieve fine dexterous manipulation is a huge leap.
BeingBeyond has released Being-H0.8, which it calls the world's first implicit tactile world action model, unifying vision, touch, action, and future state changes into a single latent space, with pretraining on more than 500,000 hours of first-person human video.
Why it matters: The model brings touch into large-scale pretraining along with a self-developed pseudo-labeling system, letting readers understand the path by which tactile data goes from a lab signal to a scalable pretraining resource.
Juwei Technology has released MicVLC, an industrial embodied brain, and proposed for the first time the VLC (Vision-Language-Calling) architecture paradigm, which decouples the cognitive layer from the execution layer, in contrast to traditional VLA models that deeply couple Action with the low-level control of specific terminal actuators.
@ChaoYuTHU has open-sourced Harness VLA, which adds a Harness Layer to organize VLA models without modifying their weights. On LIBERO-Pro, Harness VLA scores 82.4%, compared with Pi_RLinf at 50%, Berkeley RATS at 43.8%, and NVIDIA Cap-X at 18.2%. Striding AI posted this tweet and sees it as a possible sign of a shift from end-to-end VLA to Harness VLA.
Why it matters: The tweet provides comparison numbers with other VLA approaches on the same benchmark, letting readers judge how much headroom a plug-in structure that does not change weights offers.
Geek+ unveiled Gravity, its unified embodied AI framework, and Gravity 4D, its first core foundation model, at the 2026 World Artificial Intelligence Conference (WAIC), targeting long-horizon, complex physical tasks. Gravity 4D represents a fundamental leap from predicting what the world looks like to understanding how the physical world behaves, and the company also proposed its "One Core, Dual Engine" strategy.
TARS Robotics' general embodied large model AWE won the SAIL Star at WAIC 2026, becoming the only embodied intelligence foundation model to receive a SAIL Award this year. At the booth, the A1 robot used AWE to complete tasks such as packing phones into boxes, packing backpacks, and sorting small screws, and replicated a wiring harness factory's circular assembly line 1:1, with multiple A1 robots collaborating under AWE 3.5 to complete flexible wiring harness assembly.
Introducing ACT-2 Preview—the world's first robot model that can work in your home. Achieves a 99% success rate in unseen homes, operating fully autonomously. No data from you required.
Sunday Robotics previewed its robot model ACT-2, which achieved a 99.1% (±0.3%) zero-shot success rate in clothes folding across diverse unseen home environments with zero per-home adaptation, covering 9 clothing categories across 785 autonomous attempts, with an average folding quality of 4.72/5.
Why it matters: The official disclosure provides detailed quantitative data on pretraining scale and generalization gaps, allowing readers to understand how a small amount of demonstration data transfers to unseen environments.
LimX Dynamics released V³-0 alongside COSA 0.5, a humanoid VLA technology stack in its VLA-tools layer that supports mobile whole-body manipulation, running on the 43-DOF LimX humanoid robot Oli.
Why it matters: The original text explains the architectural composition of V³-0 and its human-in-the-loop real-robot reinforcement learning loop, which lets readers understand where the mobile whole-body manipulation capability comes from.
Lingchu Intelligence and the PKU-Lingchu Joint Lab released and open-sourced the dual dexterous hand general manipulation model and full-stack system EgoSteer, comprising the data processing pipeline EgoSmith, the robotics technology stack Robot-Stack, and a world-model-enhanced VLA model.
Why it matters: The original text presents a complete chain from human-hand data processing and real-robot correction to model architecture and experimental figures, allowing readers to assess its reproducibility value.
LingBot-VLA 2.0 has been open-sourced by @Robbyant, and Genrobot.AI says it provided the human data behind the model. This data includes large-scale in-the-wild egocentric video paired with hand tracking accurate to under 1 cm, processed through its Data Foundation Model.
Ant Group's Lingbo Technology has released LingBot-VA 2.0, which it describes as a world-action model pretrained from scratch for physical-world robot control, running real-time inference at 150 Hz on a single GPU and able to play air hockey, tidy desks, and pick chips from a conveyor belt.
Why it matters: It provides a same-benchmark success-rate comparison with π0.5 and single-GPU real-time inference metrics, helping to gauge where this kind of world-action model stands in terms of performance.
Ant Lingbo has open-sourced LingBot-Video, a MoE video foundation model for embodied intelligence. It uses a DiT plus sparse MoE architecture with 30B parameters and 3B activated at inference, and is claimed to be about 3 times more efficient than dense models.
Why it matters: The material provides comparison scores on RBench and the scope of the open-source release, helping readers judge the starting point and usability of this embodied video model.
LingBot-VLA 2.0 has been open-sourced. It is a next-generation embodied foundation model pretrained on 60,000 hours of real-world physical data, built on a redesigned data processing pipeline curated from 90,000 hours of robotic data and 20,000 hours of egocentric human videos.
Why it matters: The public weights and GM-100 results let readers directly compare its gap with π0.5 and GR00T N1.7 on long-horizon mobile manipulation.
BAAI has released the multimodal representation world model Wujie·RoboBrain Orca. Its core idea is to shift from predicting the next word, next frame, or next action to predicting the next world state: it first learns a unified latent world representation, then adapts to downstream tasks through three lightweight readout modules for text, images, and actions.
Why it matters: The original text explains the training paradigm shift from predicting the next output to predicting the next state, as well as the approach of a frozen backbone plus lightweight readout modules, which is transferable to readers following the world model route.
LingBot-Depth 2.0 has been released, a next-generation spatial perception model for robots that focuses on perception of glass, mirrors, and transparent objects. It was trained on 150M-scale data (50 times that of v1.0) and took first place in 12 of 16 depth completion benchmarks; on the hardest indoor scenes, depth error RMSE dropped from 0.132 to 0.062.
Why it matters: The original text provides benchmark rankings, the change in depth error on the hardest indoor scenes, and an open-weights link, helping readers judge the improvement over the previous generation and how reusable it is.
Ant Lingbo Technology has launched LingBot-Map, which uses Geometric Context Attention (GCA), an end-to-end learned streaming state, to solve the memory problem in streaming 3D reconstruction.
Ant Lingbo Technology has launched LingBot-VA, an autoregressive framework that unifies video world modeling with policy learning, jointly learning future frame prediction and action execution. Evaluation results show that the framework demonstrates clear potential in long-horizon manipulation, data-efficient post-training, and generalization to new scene configurations.
X Square Robot has released X-Tokenizer, a cross-modal embodied action tokenizer built on a semantic residual quantization (SRQ) architecture and pretrained on 17 robotic arm series and 2.4 million trajectories (including 2 billion action frames). Compared with FAST, it improves multimodal alignment by 13.5% and long-horizon task performance by 8.25%, and scores 82.8 on RoboTwin 2.0; in real-robot tests it outperforms comparison methods such as FAST and 4-level RVQ.
Why it matters: The original text presents X-Tokenizer's method design and comparative experimental data, letting readers understand the action tokenizer's multimodal alignment approach and its reusable effects.
HUMANOID has launched KinetIQ Ascend, an end-to-end visual reinforcement learning method that operates under production deployment conditions, extending its KinetIQ framework to let robots improve manipulation skills through trial and error rather than relying solely on human demonstration; the company says this is the first such method publicly demonstrated on production-line humanoid robot manipulation.
Why it matters: The original text provides specific throughput and success-rate figures for three industrial tasks and notes a pattern of predictable performance improvement with training duration, making it useful for assessing the practical value of real-world reinforcement learning on humanoid robots.
Lumos Robotics' industrial embodied brain Prime R0 ranked first worldwide by overall success rate across all single-arm fine manipulation and dual-arm collaborative manipulation tasks on MolmoSpaces, a zero-shot embodied intelligence benchmark initiated by AI2.
Galbot released AstraBrain-WBC 0.5 (paper titled Humanoid-GPT, accepted to CVPR 2026), training a Transformer whole-body controller on about 2 billion frames of retargeted motion data, achieving a 92.58% zero-shot motion tracking success rate with 0.39 ms inference latency; code and data are open-sourced.
Why it matters: This analysis clearly explains the data scale, architecture choices, and limitations of Galbot's new model, letting readers assess progress on the general motion control route.
Our Data Foundation Model (DFM) transforms raw human interactions into high-precision, multimodal, model-ready data. Through ground-truth evaluation and continuous feedback, DFM builds a scalable human data flywheel for physical AI.
Built on our Data Foundation Model (DFM), GenRobot.AI delivers higher-quality, more scalable, and more efficient Ego+Fingers data production. The industry-first DFM enables end-to-end data processing. #EmbodiedAI #HumanData #6DPose #PhysicalAI #RobotLearning #EgoData
At the inaugural EgoCross Challenge of the CVPR 2026 EgoVis Workshop, Nuoyin Intelligence took first place globally in both the Source-Limited and Open-Source tracks, becoming the only team to win championships in both tracks.
Aleph Surgical released a research preview of π0.5 driving autonomous surgical anastomosis procedures.
Kepler has released the Kepler-OmniVTLA native force-tactile omni-modal large model, paired with the Kepler-OmniTac™ force-tactile full-stack data collection solution, natively compatible with vision, six-axis force, and tactile multi-modal information.
Shima Notes talks with Wen Bowen, an NVIDIA researcher and first author of FoundationStereo and Fast FoundationStereo, about how stereo depth foundation models are changing robotic 3D perception.
Figure AI demonstrated two humanoid robots equipped with Helix-02 collaborating to tidy a bedroom within two minutes, completing tasks such as opening a door, hanging clothes, putting away headphones, closing a book, taking out trash, pushing a chair under a desk, and making a bed together. The two robots run the same learned vision-language-action (VLA) policy, with no shared planner, message passing, or central coordinator; each infers the other's intent from its actions via its own cameras. The company says that to its knowledge this is the first demonstration of multi-humanoid collaborative mobile manipulation completed by a single learned neural network directly from pixels to actions.
Why it matters: The official demonstration shows two humanoid robots driven by the same VLA system collaborating to tidy a bedroom, giving readers insight into an approach to multi-robot collaboration without a shared planner.
Genesis AI has released GENE-26.5, calling it the first "robot brain" on the path toward human-level capability. It is a full-stack approach combining a robot-native foundation model, a 1:1 humanoid robotic hand, non-invasive data-collection gloves that capture motion, force, and touch, and a simulator. The model is trained jointly on language, vision, proprioception, touch, and action, and was tested on a set of self-designed tasks; the demonstration was fully autonomous, at 1x speed, using a single model with the same weights.
Genesis AI's robotic foundation model GENE-26.5 autonomously completed a continuous one-handed fine manipulation sequence: unwrapping a straw, flipping the straw with one hand, dropping it into a cup, and then handing the drink to a person. The entire sequence was executed autonomously and filmed at 1x speed, not teleoperated. The demo showcases fluid one-handed dexterous manipulation across multiple small, precise actions.