Skip to main content
NVIDIA Introduces Cosmos Policy: World Foundation Models That Teach Robots to Act
AI·The Robot Report / NVIDIA·1 min read

NVIDIA Introduces Cosmos Policy: World Foundation Models That Teach Robots to Act

NVIDIA unveiled Cosmos Policy, a new approach to robot control that fine-tunes its Cosmos Predict world foundation model to directly generate robot actions. Instead of bolting separate perception and control modules together, Cosmos Policy encodes robot actions, physical states, and success scores as latent video frames — allowing a single model to handle visuomotor control, future-state prediction, and action planning simultaneously. The system achieved state-of-the-art results on LIBERO and RoboCasa manipulation benchmarks, outperforming both diffusion policies trained from scratch and vision-language-action models. By inheriting the pretrained model's understanding of physics and temporal dynamics, Cosmos Policy represents a significant step toward general-purpose robot intelligence built on video foundation models.

NVIDIACosmosWorld Foundation ModelsAIManipulationRobot Control
Read original source

Related News

View All News →
No image available
AI

Vision-Language-Action Models Are Replacing Modular Robotics Pipelines

A new class of AI models called vision-language-action (VLA) systems is changing how humanoid robots process instructions and interact with the world. Instead of separate modules for seeing, planning, and moving, VLAs like Figure AI's Helix, NVIDIA's GR00T N1, and Google DeepMind's RT-1 combine vision, language understanding, and motor control into a single end-to-end model. Helix uses a dual-system design — a large vision-language backbone handles high-level reasoning while a smaller action model runs at high frequency for precise manipulation. GR00T N1 takes a generalist approach, training across multiple robot types and tasks. The practical result: robots that can follow natural language instructions, carry out multi-step tasks, and adapt to new environments without hand-tuned pipelines. Several teams have demonstrated on-device deployment with reduced latency. The shift matters because it moves humanoid robots closer to general-purpose assistants that understand what you ask and can actually do it.

·The Robot Report
No image available
Humanoid

Humanoid Launches KinetIQ: An AI Framework for Orchestrating Fleets of Humanoid Robots

UK-based Humanoid unveiled KinetIQ, an AI framework for end-to-end orchestration of humanoid robot fleets across both wheeled and bipedal platforms. The system coordinates fleet-level operations and individual robot behavior through four cognitive layers — from high-level task allocation and workflow optimization down to task execution using Vision-Language-Action models and whole-body control trained via reinforcement learning. The announcement comes after Humanoid completed a live logistics trial at Siemens processing 60 totes per hour, and secured a deal with Schaeffler to deploy hundreds of its robots across factories. Founded in 2024, Humanoid has quickly positioned itself as a leading European contender in the commercial humanoid race.

·IEEE Spectrum / Humanoid
Figure Unveils Helix 02: Full-Body Neural Network Controls Humanoid Walking, Manipulation, and BalanceHumanoid

Figure Unveils Helix 02: Full-Body Neural Network Controls Humanoid Walking, Manipulation, and Balance

Figure AI announced Helix 02, a single neural network that controls its humanoid robot's entire body directly from camera pixels. The system demonstrated a continuous 4-minute autonomous kitchen task — unloading a dishwasher, walking across the room, stacking items in cabinets, and reloading the dishwasher — with no human intervention. Helix 02 replaces over 109,000 lines of hand-engineered C++ with a unified visuomotor network trained on 1,000+ hours of human motion data.

·Figure AI