Articles tagged with "multi-agent-systems"
VERSES multi-agent robotics model works without pre-training - The Robot Report
VERSES AI Inc. has developed a novel multi-agent robotics architecture based on hierarchical active inference that enables robots to perform typical household tasks more effectively than existing models without requiring any pre-training. Unlike traditional robotics approaches—drive-by-wire systems that rely on pre-programming and deep learning models that need extensive training data—VERSES’ system adapts dynamically by exploring its environment, using integrated vision, planning, and control modules. This approach allows robots to handle unexpected obstacles and changes in their surroundings, overcoming common limitations such as freezing or halting when encountering unfamiliar situations. The company, founded in 2020 and based in Vancouver, emphasizes that its platform is inspired by principles from science, physics, and biology to generate reliable predictions and decisions under uncertainty. In comparative tests involving household tasks like tidying a room, preparing groceries, and setting a table, the VERSES model achieved a 66.5% success rate, outperforming a deep learning baseline that scored 54.7%. VERSES claims this
roboticsartificial-intelligencemulti-agent-systemsadaptive-robotsautomationVERSES-AIrobotics-architectureTiny but mighty: This AI mini-model outsmarted Microsoft on Meta’s GAIA benchmark
Coral Protocol, a London-based AI company, has achieved a significant milestone by developing a multi-agent AI "mini-model" system that outperformed Microsoft’s agent platform by approximately 34% on Meta’s GAIA benchmark. GAIA is a challenging test suite comprising nearly 450 complex real-world tasks requiring reasoning, web browsing, data analysis, and tool use. While human participants typically answer about 92% of GAIA questions correctly, advanced large models like GPT-4 manage only around 15%. Coral’s mini-model scored the highest among small-scale AI systems, surpassing Microsoft-backed Magnetic-UI, which scored about 30%. Coral’s approach diverges from the traditional AI scaling method of building massive models with billions of parameters. Instead, it employs horizontal scaling by orchestrating many specialized, lightweight mini-models that collaborate in real time, each excelling at specific tasks such as natural language understanding or coding. This collective intelligence framework enables faster, more cost-effective, and potentially more secure
IoTartificial-intelligenceAI-assistantsmulti-agent-systemsAI-mini-modelshorizontal-scalingCoral-ProtocolShlomo Zilberstein wins the 2025 ACM/SIGAI Autonomous Agents Research Award
robotautonomous-agentsmulti-agent-systemsdecision-makingreinforcement-learningresearch-awardAIMulti-agent path finding in continuous environments
robotautonomous-drivingmulti-agent-systemspath-findingwarehouse-logisticscollision-avoidancerobotics