AMD Demonstrates an End-to-End Embodied-AI Stack
An AMD-accelerated stack spans data-center training, simulation and Ryzen AI edge deployment for VLA manipulation.
Event Type: research publication
World-model research identifies limits in learning hidden physical parameters
A controlled study found prediction objectives determine whether robot world models retain mass, stiffness, drag, and force information.
SymmGrid accelerates direct on-robot learning for contact tasks
Symmetry-based trajectory augmentation improved training speed and success rates on real robot insertion, routing, and relocation tasks.
HERO demonstrates self-improving manipulation without human demonstrations
Researchers introduced HERO, a hierarchical embodied agent that autonomously gathers experience and consolidates reusable manipulation policies.
MulRobBench Extends Safety Evaluation to Multi-Robot Decisions
A benchmark evaluates decision-level safety and reliability across multi-robot and aerial scenarios.
Spatial Intelligence Is Decomposed for Physical AI Evaluation
Researchers propose more granular evaluation of spatial reasoning needed for physical manipulation and navigation.
Trustworthy Embodied Intelligence Gets a Graded Systems Framework
A systems framework links robust task execution, hardware safeguards, evidence and deployment governance.
World Models Receive a Security Threat Model
Researchers map attacks across sensing, world modelling, planning, execution, feedback and long-term adaptation.
Humanoid social memory becomes lightweight and persistent
A lightweight episodic memory module enables a humanoid robot head to retain personalized conversational context across sessions.
General multimodal model shows emergent embodied reasoning
An exploratory mobile-robot study reports spatial memory and action consequence reasoning from a general multimodal model.
