Tag: Vision language action

NVIDIA Cosmos 3 Launches as World Action Model to Replace VLA Paradigm

NVIDIA unveiled Cosmos 3 at GTC Taipei -- an open-weights physical AI model with native robotic action generation. The model displaces the VLA paradigm in favor of WAM, ranking first on four open leaderboards.

Yann LeCun: LLMs Are a Dead End. Paradigm Shift by 2027

Yann LeCun predicts a robotics paradigm shift by 2027 and criticizes VLA models as brittle and data-inefficient. AMI Labs is betting on JEPA architecture and LeWorldModel as alternatives to LLM-based approaches.

Man vs. Machine: Figure AI Intern Beats Humanoid Fleet in 10-Hour Sorting Race

Figure AI staged a 10-hour head-to-head between a human intern and its autonomous Figure 03 fleet. Aime the intern won narrowly — 12,924 packages to 12,732 — but the robot fleet has since continued working uninterrupted for over 116 hours.

NVIDIA Maps Robotics' Future on the LLM Evolution Path

NVIDIA's Jim Fan at AI Ascent 2026 outlined a detailed robotics development roadmap based on replicating four stages of large language model evolution — from pre-training to autonomous research.

Tutor Intelligence launches Data Factory: 100 robots learning from humans in real time

MIT startup Tutor Intelligence launched DF1 — the US's largest robotic data factory with 100 Sonny robots. The company detects behavior errors 100x faster and is already commercializing Cassie mobile manipulator with usage-based pricing.

Physical AI brings new governance demands for autonomous systems

As physical AI systems scale across factories and infrastructure, a critical governance gap is emerging: who defines the boundaries of AI that can issue instructions to machines, and how? The industry is still working out the answers.

Figure AI Opens Its Campus: Never Fall Protocol, Offline AI, and a $500/Month Home Robot Lease

Figure AI has opened its Sunnyvale campus to journalists, revealing the Never Fall resilience protocol, 2kW wireless charging, and a $400–$600/month home robot lease model. CEO Brett Adcock detailed the Vulcan AI system and the planned Figure 4.

Meta Acquires Assured Robot Intelligence to Build the Android of Humanoid Robotics

Meta Platforms has acquired ARI, a startup building AI models for humanoid robots. The team joins Meta Superintelligence Labs as the company bids to become the foundational hardware and software platform for the entire robotics industry.

World Action Models: What They Are and How They Work

Vision-Language-Action (VLA) models have become the dominant approach to building AI-driven robot control systems over the past several years. Their emerging successor — the World Action Model (WAM) — represents a distinct architectural category that replaces direct image-to-action mapping with video generation as an intermediate planning mechanism. DreamZero, developed by a team at NVIDIA and published in February 2026 as a research paper on arXiv, is the first publicly described system of this class to operate in real time on a real robot. Understanding how it works matters because it points toward a plausible direction for the next generation of robotic foundation models.

ShengShu unveils Motubrain: one AI brain for every robot platform

ShengShu Technology has unveiled Motubrain, a World Action Model that unifies video generation, world modeling, and action control in a single Mixture-of-Transformers architecture. Backed by a $293M Series B from Alibaba Cloud, the model tops RoboTwin 2.0 with 96% success across 50 randomized tasks.

AI² Robotics otwiera platformę dla deweloperów robotów humanoidalnych. Jaki jest status tej chińskiej firmy?

Japonia stawia na „Physical AI”: Powstaną globalne centra badawcze robotyki nowej generacji

Otwarta rewolucja w robotyce: Cztery siły walczą o prymat nad „cyfrowym mózgiem”