robotics

NVIDIA has added four open-weight checkpoints to its Cosmos 3 family: Transfer-Auto, Sim-Auto, Sim-Transfer and Sim-Bimanual. Each is a post-trained Cosmos 3 Nano with 16B trainable parameters, released under OpenMDW1.1 for commercial and non-commercial use.

China Merchants Lion Rock AI Lab released Lion-ER-35B-A3B, a mixture-of-experts embodied reasoning vision-language model with about 35B total and about 3B active parameters per token. The lab says it beats its base model on all ten public benchmarks tested, with an equal-weighted average of 65.29. Weights are open on Hugging Face.

Ant Group's embodied AI unit says its open-source LingBot robot foundation models have passed 2.2 million downloads and 33,000 GitHub stars in nine months. It also presented a 2.0 lineup and a robotic pharmacy already operating in several Guoda Pharmacy stores in Shanghai.

Xiaomi says its humanoid robot worked an eight-hour shift at its car factory, covering the full daily production demand of the center console side-cover sorting station. The company also started a home-service trial of its Xiaomi-Robotics-1 embodied foundation model at a Xiaomi Youth Apartment.

Google's updated Genie 4 points to new uses for world models in robotics and AI testing, according to AI Business. The outlet says enterprises must weigh that potential against accuracy gaps.

ACE Robotics and Nanyang Technological University's S-Lab have unveiled Mundo, a world model that builds an interactive, physics-simulated scene from one to 12 photos of a real place. Code and weights are due by the end of November, according to the GitHub repository. The team lists reconstruction times of 10 to 20 minutes and rendering at about 1.5 frames per second on eight A800 GPUs.

AgiBot (智元) released two models at its Innovation Day: GO-3 Preview, which it says ranks first on three embodied intelligence benchmarks, and WITA-OMNI, which it says ranks first on four. It also introduced the Worken metric, the AHA3 hardware architecture, the G2 Max robot and Lingqu OS 1.0.

AgiBot launched GO-3 Preview on October 29, an embodied AI model built on a World Reasoning Action Model (WRAM) architecture. The company says it ranks first on RoboDojo, RoboTwin and RoboColiseum, and plans to open-source the full model by the end of 2026.

AGIBOT introduced Worken, a proposed framework for measuring how much productive work robots perform once deployed, at AGIBOT Innovation Day 2026 in Shanghai on October 29, 2026. The company also debuted the GO-3 Preview embodied foundation model and the WITA-Omni interaction model.

A preview of AgiBot's next robot manipulation foundation model, GO-3 Nano Preview, ranks first on RoboDojo's simulation leaderboard with a 61.90 average score and a 56.42 percent average success rate. It leads three of six dimensions and trails other entries on memory and generalisation.

iFLYTEK detailed SparkAudio 1.0 at its 2026 Global 1024 Developer Festival. The speech foundation model combines a 0.65-billion-parameter audio encoder with a 30B-A3B mixture-of-experts language model, trained on 13 million hours of audio. iFLYTEK also launched Robot Super Brain Platform 3.0 for embodied robots.

Primotion, a Shenzhen startup founded in July 2026 to build embodied-AI chips, has unveiled Paimon-0, an embodied foundation model. The company says it cuts single action-chunk inference from about 3,000 ms to 300 ms and ranks first on several public robot benchmarks. All results are self-reported.

Amap, part of Alibaba Group, has released λ0, which it describes as the first whole-body foundation model based on open-world human behavior. According to Amap, it learns movement, posture and two-handed manipulation from human data alone, cuts collection time per demonstration from about 100 seconds to 29, and raises real-robot task success 22.5% over the strongest baseline.

Robot-brain developer Qianjue launched Polibrain, a world model that predicts task-relevant state changes instead of future video frames. It is adapted to quadruped, humanoid and wheeled robots, according to Pandaily.

SenseTime has open-sourced the SenseNova-RoboRSI evaluation harness and a 27B embodied model. The release follows a study of GPT-6-Astra across 12 simulated benchmarks and 8 real robots. The figures are SenseTime's own.

On October 21, Shengshu Technology's world model team began a public livestream of the full Motubrain 2.0 training process, planned to run about half a month. The embodied model is due to be open-sourced, and the stream covers pretraining, post-training, agent task execution and recursive self-improvement (RSI) exploration.

Elorian, the Palo Alto startup that raised a $55M seed at a $300M valuation, has released Glimpse as a research preview with API access at $0.60 per 1M input tokens and $3 per 1M output tokens. The model can crop, pull video frames, and run code mid-reasoning. Benchmark claims come from Elorian's own evaluations.

Google DeepMind is positioning Gemini 4 Argon as one model that controls robots, builds simulated training environments, and grades training data. Google claims 87% on the ERQA benchmark, ahead of Claude Opus 5.5 at 82.3%. Artificial Analysis's AA-Robotics test, published a day later, placed Argon third at 55 out of 100.

Google DeepMind says Gemini 4 Argon scores 87% on the Embodied Reasoning (ERQA) benchmark using pure reasoning, ahead of every other frontier model. A new robotics post demonstrates robot control, Real2Sim reconstruction, physical understanding, and self-improvement workflows. Pricing, context window, and availability were not disclosed.

Odyssey has opened a free public research preview of Odyssey-3, an autoregressive diffusion transformer that generates interactive environments from text prompts in real time. The base model has 14 billion parameters and outputs 832 × 480 video. Odyssey's headline Physics-IQ score of 66.1 does not meet the benchmark's own requirements for a record claim.

Reka AI has released a research preview of Rho-1, a 19-billion-parameter omni-model that processes and generates text, images, video, and robot control actions in a single network. Reka says it uses no tool calls or external models. Context window, pricing, and benchmark scores have not been disclosed.

Researchers built HomeBody, a system that connects GPT-6 Astra directly to a Unitree G1 robot's skill library, letting it explore, map, and tidy an unfamiliar kitchen without a trained control layer in between. The team reports latency, overheating servos, and compute cost as current limitations.

Black Forest Labs has released FLUX 3 Action, an open-weight robotics model built on its FLUX 3 multimodal foundation. The 7-billion-parameter model reads multi-camera video feeds and predicts what a robot should do next, claiming a record success rate on the RoboLab-120 leaderboard while running nearly 4x faster than the previous best open model.

Anthropic is setting up a physical biology lab in the San Francisco area where Claude will guide robots through drug experiments with minimal human involvement. The move follows the company's $400 million acquisition of Coefficient Bio and the addition of Novartis CEO Vas Narasimhan to its board.

Runway has shared research into generating video frame by frame in real time, letting users steer output as it streams rather than waiting for a finished clip. The approach builds on GWM-1, Runway's General World Model introduced in December 2025, and no availability timeline has been announced.

A new benchmark called RoboHarm tested whether AI models controlling robotic arms would refuse dangerous commands. GPT-6 Astra completed 60 of 100 dangerous tasks and Claude Fable 5.1 completed 34, with neither model showing a reliable safety layer.

A new robotics benchmark called StationeryBench shows OpenAI's GPT-6 Astra completing 7 of 100 desk-object manipulation tasks versus zero for Ai2's MolmoAct2, with a median progress score of 46 against 12. Cornell/DeepMind researcher Yoav Artzi calls the result a 'step change in spatial reasoning.'

Anthropic has released a research preview of the Model Hardware Standard (MHS), a protocol that lets AI agents discover and control physical devices like robotic arms and liquid handlers through a single interface. Built with HHMI Janelia Research Campus, the spec has been tested by Genentech, Carnegie Mellon, and QuEra, with Anthropic claiming it cuts hardware integration time from weeks to hours.

Robotics startup Generalist AI has released GEN-1.5, a model that loads a short video demonstration into its context window and performs the task without additional training. The company reports a 59 percent success rate zero-shot and 83 percent after light fine-tuning, though all results are self-reported.

A new Epoch/METR benchmark called MirrorCode shows Claude Opus 4.7 reimplementing large software programs from scratch in tasks estimated to take humans 2-17 weeks, for $251 in inference cost. Separately, Anthropic reports Opus 4.7 completed a suite of quadruped robot tasks in 9 minutes 35 seconds, down from 181 minutes with an earlier model assisting humans.

Black Forest Labs has released Flux 3, a multimodal foundation model trained jointly on images, video, and audio that generates videos up to 20 seconds long with synchronized native audio. The company also introduced Flux-mimic, a robotics action model already being tested at Audi.

NVIDIA has released Cosmos 3 Edge, a 4-billion-parameter open world model designed for edge AI systems. The model delivers real-time robot control at 15 Hz on NVIDIA Jetson devices, generating 32 actions per inference at 640×360 resolution.

Nvidia released Cosmos 3 Edge, a world model designed for robots and vision AI agents to perceive and navigate physical environments in real time. The company announced partnerships with Japanese industrial giants including Fujitsu, Hitachi, and Kawasaki Heavy Industries as part of its physical AI expansion.

Mistral AI released Robostral Navigate, an 8B parameter model that enables autonomous robot navigation using only a single RGB camera. The model achieves 76.6% success on R2R-CE validation unseen benchmarks, outperforming multi-sensor approaches by 4.5 percentage points despite using no depth sensors or LiDAR.

NVIDIA released Cosmos3-Super-Text2Image, a 64-billion parameter text-to-image generation model as part of its Cosmos3 collection of omnimodal world models. The model uses a Mixture-of-Transformers architecture combining autoregressive and diffusion transformers, designed for Physical AI applications including robotics and autonomous vehicles.

NVIDIA released Cosmos 3, an omnimodal world foundation model platform for Physical AI spanning robotics, autonomous driving, and industrial environments. The flagship Cosmos3-Super variant contains 64 billion parameters and generates video, images, audio, and action commands from text, image, video, and action trajectory inputs using a Mixture-of-Transformers architecture.

NVIDIA released Cosmos3-Super, a 64-billion parameter omnimodal foundation model that generates video, images, audio, and action commands from combinations of text, image, video, and action trajectory inputs. The model, part of the Cosmos3 collection, targets Physical AI applications including robotics, autonomous vehicles, and industrial automation.

NVIDIA has released Cosmos3-Nano, a 16-billion parameter omnimodal world model capable of generating video, audio, images, and robot action commands from combinations of text, image, video, and action trajectory inputs. The model supports a 256K token context window and is designed for Physical AI applications including robotics, autonomous vehicles, and smart manufacturing environments.

NVIDIA has released Cosmos 3, a unified omni-model that combines world generation, physical reasoning, and action generation in a single architecture. Available in 8B (Nano) and 32B (Super) parameter versions on Hugging Face, Cosmos 3 uses a Mixture-of-Transformers architecture to process text, image, video, audio, and action modalities without switching between separate models.

NVIDIA released LocateAnything-3B, a 3-billion parameter vision-language model that predicts bounding boxes in parallel rather than token-by-token, achieving up to 2.5× higher throughput compared to autoregressive approaches. The model, trained on 12M images with 138M+ queries and 785M bounding boxes, supports object detection, GUI element grounding, and robotics perception.

Google DeepMind has integrated its Genie world model with Street View's 280 billion images spanning 110 countries, enabling users to explore AI-generated simulations of real locations. Waymo is already using Genie 3 to train self-driving cars on rare scenarios like tornadoes and unexpected obstacles.

Google DeepMind launched Street View integration with its Genie 3 world model at I/O 2026, allowing users to simulate real-world locations from 280 billion images across 110 countries. The feature enables environment modification including weather changes and supports robotics training, with initial access for U.S. Ultra subscribers expanding globally.

NVIDIA published a technical guide for parameter-efficient fine-tuning of its Cosmos Predict 2.5 world model using LoRA and DoRA adapters. The method allows teams to adapt the 2B-parameter model to robot manipulation tasks on a single 80GB GPU, generating synthetic training trajectories from just 92 demonstration videos.

NVIDIA released GR00T N1.7, a 3-billion parameter open-source Vision-Language-Action model for humanoid robots with commercial licensing. The model was trained on 20,854 hours of human egocentric video data and demonstrates the first documented scaling law for robot dexterity, where increasing human video data from 1,000 to 20,000 hours more than doubles task completion rates.

Physical Intelligence published research showing its π0.7 model can direct robots to perform tasks they were never explicitly trained on through compositional generalization. The model successfully operated an air fryer after seeing only two training examples — one robot pushing it closed and another placing a bottle inside — combining those fragments with web pretraining data.

AI2 has released MolmoSpaces and MolmoBot, robotics models trained exclusively in simulation that transfer directly to real robots without manual real-world data collection or fine-tuning. The approach eliminates months of teleoperated demonstrations typically required for simulation-trained robots. Both systems are open-source.

AI2 is developing physical AI systems trained primarily on virtual simulation data rather than expensive real-world demonstrations. The approach, demonstrated through projects like MolmoBot, addresses the historical bottleneck of manually collecting hardware training data.

ABB and NVIDIA have partnered to deploy physical AI simulation in factory automation, addressing the critical sim-to-real gap that has limited intelligent robotics deployment. The approach uses digital physics simulation to train models that transfer reliably to actual factory floors, reducing production hurdles and securing measurable ROI.

World Labs, the AI startup founded by Fei-Fei Li, has raised $1 billion in new funding to develop spatial intelligence—AI systems capable of understanding and operating in three-dimensional physical environments. The capital will fund the development of world models, a class of AI architecture designed to reason about spatial relationships and physical interactions.