Articles tagged with "AI-models"
Nvidia launches Alpamayo, open AI models that allow autonomous vehicles to ‘think like a human’
Nvidia has introduced Alpamayo, a new suite of open-source AI models, simulation tools, and datasets aimed at advancing autonomous vehicle (AV) capabilities by enabling them to reason through complex driving scenarios like humans. Central to this release is Alpamayo 1, a 10-billion-parameter vision language action (VLA) model that employs chain-of-thought reasoning to break down problems step-by-step and select the safest driving actions, even in rare or unfamiliar situations such as traffic light outages. This model’s code is publicly available on Hugging Face, allowing developers to fine-tune it for various applications, including simpler driving systems, auto-labeling video data, and decision evaluators. Nvidia also encourages combining real and synthetic data generated via its Cosmos platform to enhance training and testing. Alongside Alpamayo 1, Nvidia is releasing an extensive open dataset comprising over 1,700 hours of driving data from diverse geographies and conditions, focusing on rare and complex scenarios. To support
robotautonomous-vehiclesAI-modelssimulation-toolsNvidiaopen-source-AIphysical-robotsPeripheral Labs taps into self-driving car sensors to bring sports fans right into the game
Peripheral Labs, a Canadian startup founded in 2024 by Kelvin Cui and Mustafa Khan, is leveraging technology originally developed for self-driving cars to revolutionize sports viewing through volumetric video generation. This technology uses multiple cameras to capture 3D footage, allowing fans to watch plays from various angles and interact with the video in immersive ways, such as tracking specific players or freezing moments for detailed analysis. By applying robotics perception and 3D vision expertise from autonomous vehicle research, the company has reduced the number of required cameras from over 100 to as few as 32, significantly lowering costs and operational complexity for sports leagues and broadcasters. The platform also integrates biomechanical data by using sensor technology similar to that in driverless cars, enabling detailed analysis of player movements, such as joint flexion, which can assist coaches in improving player performance. Peripheral Labs aims to make this advanced technology affordable and scalable, targeting multi-year contracts with teams and broadcasters. The startup recently raised $3.6 million in seed funding
robotcomputer-vision3D-reconstructionself-driving-car-sensorssports-technologyAI-modelsvolumetric-video-generationNvidia bulks up open source offerings with an acquisition and new open AI models
Nvidia is strengthening its presence in open source AI through two major initiatives: the acquisition of SchedMD and the release of a new family of open AI models. SchedMD, founded in 2010 by the original developers of the widely used open source workload management system Slurm, has been a long-term partner of Nvidia. The acquisition, with undisclosed terms, aims to leverage SchedMD’s technology as critical infrastructure for generative AI, enabling Nvidia to accelerate access to diverse computing systems. Nvidia plans to continue investing in this technology to support AI development at scale. In addition to the acquisition, Nvidia introduced the Nemotron family of open AI models, which it claims to be the most efficient open models for building accurate AI agents. This lineup includes the Nemotron 3 Nano for targeted tasks, Nemotron 3 Super for multi-agent AI applications, and Nemotron 3 Ultra for more complex tasks. Nvidia’s CEO Jensen Huang emphasized that Nemotron represents a move toward open innovation,
robotAI-modelsNvidiaopen-source-AIgenerative-AIworkload-managementGPUsFrontier supercomputer powers first nuclear-specific AI for reactors
The article discusses the development of the first nuclear-specific artificial intelligence (AI) models powered by the Frontier supercomputer, the world’s second-fastest and first exascale machine, located at Oak Ridge National Laboratory. Developed by tech startup Atomic Canyon, the Neutron platform leverages Frontier’s immense computational power to train AI capable of accurately searching and interpreting complex nuclear industry documents, including regulations, maintenance logs, and engineering records. This innovation aims to drastically reduce the time, labor, and resources traditionally spent on navigating vast nuclear documentation, which can consume tens of thousands of staff hours annually at plants like California’s Diablo Canyon. The initiative originated from the challenges faced by Diablo Canyon, California’s only operational nuclear plant, which had to extend its license to 2030 and manage a massive regulatory application under tight deadlines. Existing commercial AI tools lacked the precision needed to handle nuclear-specific jargon and abbreviations, prompting Atomic Canyon to build a dedicated AI model from scratch. Using 20,000 GPU hours on Frontier
energyartificial-intelligencenuclear-powersupercomputerAI-modelsenergy-industrynuclear-reactorsNVIDIA debuts first open reasoning AI for self-driving vehicles
NVIDIA has introduced a suite of open-source AI models and tools aimed at advancing autonomous vehicles, robotics, and speech processing. Central to this launch is Alpamayo-R1 (AR1), the world’s first open reasoning vision-language-action (VLA) model designed for self-driving cars. AR1 integrates chain-of-thought reasoning with path planning to navigate complex driving scenarios by evaluating possible trajectories and contextual data, enabling human-like decision-making in challenging environments such as crowded intersections or lane closures. Built on NVIDIA’s Cosmos Reason platform, AR1 is available for customization by researchers for non-commercial use and has demonstrated improved reasoning capabilities through reinforcement learning post-training. Beyond AR1, NVIDIA’s Cosmos platform offers additional tools for physical AI development, including LidarGen for generating lidar data, Omniverse NuRec Fixer for neural reconstruction cleanup, Cosmos Policy for robot behavior creation, and ProtoMotions3 for training humanoid robots in simulated settings. These resources are already being utilized by ecosystem partners and academic
robotautonomous-vehiclesAI-modelsNVIDIA-Cosmosreinforcement-learninglidar-simulationhumanoid-robotsNvidia announces new open AI models and tools for autonomous driving research
Nvidia has unveiled new AI infrastructure and models aimed at advancing physical AI applications, particularly in robotics and autonomous vehicles. At the NeurIPS AI conference, the company introduced Alpamayo-R1, described as the first vision-language-action model specifically designed for autonomous driving research. This model integrates visual and textual data to enable vehicles to perceive their environment and make informed decisions, leveraging Nvidia’s existing Cosmos reasoning model family, which was initially launched in January 2025. Alpamayo-R1 is intended to help autonomous vehicles achieve level 4 autonomy—full self-driving capability within defined areas and conditions—by providing them with “common sense” reasoning to handle complex driving scenarios more like humans. In addition to the new model, Nvidia released the Cosmos Cookbook on GitHub, a comprehensive resource including step-by-step guides, inference tools, and post-training workflows to assist developers in customizing and training Cosmos models for various applications. This toolkit covers essential processes such as data curation, synthetic data generation, and model
robotautonomous-vehiclesAI-modelsNvidiaphysical-AIautonomous-drivingvision-language-modelsGoogle's Gemini model lets humanoid robot carry out multimodal tasks
Google DeepMind has unveiled advancements in its humanoid robots powered by the Gemini Robotics 1.5 AI models, enabling them to perform complex, multi-step tasks through multimodal reasoning. Demonstrated in a recent video, the bi-arm Franka robot successfully completed the "banana test," sorting different fruits by color into separate plates, showcasing improved capabilities over previous models that could only follow single-step instructions. Another test featured Apptronik’s Apollo humanoid sorting laundry by color, even adapting to changes in basket positions mid-task, highlighting the robots' enhanced perception and adaptability. The Gemini Robotics 1.5 family includes two complementary models: one that converts visual inputs and instructions into actions, and another that reasons about the environment to create step-by-step plans. This agentic framework allows robots to autonomously study their surroundings, make decisions, and execute tasks such as sorting waste according to local recycling rules by researching guidelines online and applying them in real time. Google emphasizes safety in these models, incorporating risk assessment
roboticshumanoid-robotsAI-modelsmultimodal-tasksautonomous-robotsrobot-perceptionrobot-reasoningFieldAI raises $405M to build universal robot brains
FieldAI, a robotics AI company, announced a $405 million funding raise to develop universal "robot brains" capable of controlling diverse physical robots across varied real-world environments. The latest funding round, including a $314 million tranche co-led by Bezos Expedition, Prysm, and Temasek, adds to backing from investors such as Khosla Ventures and Intel Capital. FieldAI’s core innovation lies in its "Field Foundation Models," which integrate physics-based understanding into embodied AI—AI that governs robots physically navigating environments—enabling robots to quickly learn, adapt, and manage risk and safety in new settings. This physics-informed approach contrasts with traditional AI models that often lack risk awareness, making FieldAI’s robots better suited for complex and potentially hazardous environments. Founder and CEO Ali Agha emphasized that their goal is to create a single, general-purpose robot brain that can operate across different robot types and tasks, with a built-in confidence measure to assess decision reliability and manage safety thresholds. Agha’s decades
robotartificial-intelligenceembodied-AIrobotics-safetyrobot-learningAI-modelsrobotics-technologyBuzzy AI startup Multiverse creates two of the smallest high-performing models ever
Multiverse Computing, a leading European AI startup based in Spain, has developed two of the smallest yet high-performing AI models, humorously named after animal brain sizes: SuperFly and ChickBrain. These models are designed to be embedded in Internet of Things (IoT) devices and run locally on smartphones, tablets, and PCs without requiring an internet connection. SuperFly, inspired by a fly’s brain, is a compressed version of Hugging Face’s SmolLM2 135 model with 94 million parameters, optimized for limited data and voice-command applications in home appliances. ChickBrain, with 3.2 billion parameters, is a compressed version of Meta’s Llama 3.1 8B model and offers advanced reasoning capabilities, outperforming the original in several benchmarks such as MMLU-Pro, Math 500, GSM8K, and GPQA Diamond. The key technology behind these models is Multiverse’s proprietary quantum-inspired compression algorithm called CompactifAI, which significantly reduces model
IoTAI-modelsmodel-compressionedge-computingembedded-AIquantum-inspired-algorithmssmart-devicesNvidia unveils new Cosmos world models, infra for robotics and physical uses
Nvidia has introduced a suite of new AI models and infrastructure aimed at advancing robotics and physical AI applications. The centerpiece is Cosmos Reason, a 7-billion-parameter vision-language model designed to enable robots and AI agents to "reason" by leveraging memory and physics understanding. This capability allows Cosmos Reason to function as a planning model, helping embodied agents determine their next steps, and it can be applied in data curation, robot planning, and video analytics. Alongside Cosmos Reason, Nvidia unveiled Cosmos Transfer-2, which accelerates synthetic data generation from 3D simulations or spatial controls, and a faster, distilled version of Cosmos Transfers optimized for speed. In addition to the AI models, Nvidia announced new neural reconstruction libraries that facilitate 3D simulation of real-world environments using sensor data, with integration into the open-source CARLA simulator. The company also updated its Omniverse software development kit and introduced new hardware solutions, including the Nvidia RTX Pro Blackwell Servers tailored for robotics workflows and the DG
roboticsAI-modelsNvidia-Cosmossynthetic-data-generation3D-simulationrobot-planningneural-reconstructionMeta’s new AI helps robots learn real-world logic from raw video
Meta has introduced V-JEPA 2, an advanced AI model trained solely on raw video data to help robots and AI agents better understand and predict physical interactions in the real world. Unlike traditional AI systems that rely on large labeled datasets, V-JEPA 2 operates in a simplified latent space, enabling faster and more adaptable simulations of physical reality. The model learns cause-and-effect relationships such as gravity, motion, and object permanence by analyzing how people and objects interact in videos, allowing it to generalize across diverse contexts without extensive annotations. Meta views this development as a significant step toward artificial general intelligence (AGI), aiming to create AI systems capable of thinking before acting. In practical applications, Meta has tested V-JEPA 2 on lab-based robots, which successfully performed tasks like picking up unfamiliar objects and navigating new environments, demonstrating improved adaptability in unpredictable real-world settings. The company envisions broad use cases for autonomous machines—including delivery robots and self-driving cars—that require quick interpretation of physical surroundings and real
roboticsartificial-intelligencemachine-learningautonomous-robotsvideo-based-learningphysical-world-simulationAI-modelsMeta says its Llama AI models have been downloaded 1.2B times
MetaLlama-AIartificial-intelligencedownloadstechnology-newsmachine-learningAI-modelsAlibaba unveils Qwen 3, a family of ‘hybrid’ AI reasoning models
AlibabaQwen-3AI-modelshybrid-AImachine-learningtech-newsopen-source-AI