Embodied Reasoning Expands to Longer Tasks
Google separated higher-level planning from physical control while broadening multi-step task execution.
Semiconductor Domain: AI accelerators
Action-Conditioned World Models Move Into Planning
Action-conditioned predictive models are used to improve generalizable embodied decision-making.
HERO demonstrates self-improving manipulation without human demonstrations
Researchers introduced HERO, a hierarchical embodied agent that autonomously gathers experience and consolidates reusable manipulation policies.
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.
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.
Japan’s Noetra positions sovereign physical AI as national infrastructure
Government-backed Noetra plans national AI infrastructure and a Japanese physical-AI model supported by major industrial groups.
Noetra begins full-scale development of a Japan-made multimodal foundation model
Noetra, Sony, SoftBank, NEC, Honda, AIST, and partners started full-scale development of a sovereign multimodal model.
AMD embedded processors enter a new humanoid platform program
Foundation selected AMD Ryzen AI Embedded processors for its next humanoid platform, linking embedded compute directly to industrial deployment.
Qualcomm presents a task-oriented humanoid stack in India
Qualcomm demonstrated a robotics stack in India targeting logistics, manufacturing, and retail through task-specific embodied-AI capabilities.
China’s Robot Supply Chain Is Assessed as Deep but Still Compute-Dependent
MERICS assessed China’s broad robot manufacturing base while noting continuing dependence on foreign AI compute and software ecosystems.
