SenseTime's SenseNova division is the institutional home of the Kairos physical AI world model project.
“GPT-6-Astra outperforms existing methods by substantial margins, including recent agentic approaches that also build on GPT-6-Astra.”
Source→“A single, general-purpose model (GPT-6-Astra) can outperform specialized Vision-Language-Action (VLA) models across a wide range of robotic tasks without updating its weights.”
Source→“On a real Unitree G1 with Wuji hands, it commands arm-joint targets and binary hand open/close states for manipulation (2/3), and turns and walks through the vendor velocity interface (3/3)”
Source→“We built a fully interactive one-bedroom apartment in NVIDIA Isaac Sim [114] through a reproducible pipeline”
Source→“it controls an AgiBot G2 [167] through the robot's own interface”
Source→“models focusing on static spatial intelligence, such as World Labs' Marble and TeleWorld, excel at building explorable 3D worlds.”
Source→“By anchoring previous physical and behavioral priors into robot-specific interaction data (e.g., AgiBotWorld-Beta, Droid), Kairos achieves perception-action grounding.”
Source→“Public sources include...specialized corpora for robotics (e.g., AgiBotWorld-Beta, Droid)”
Source→“The GitHub repository is listed as `kairos-agi/kairos-sensenova`, indicating SenseTime's SenseNova division is the institutional home of the Kairos project.”
Source→“Fei Wang — Listed as first named author on the Kairos team.”
Source→“Shan You — Second named author. Co-leads the Kairos team alongside Fei Wang.”
Source→“Participants: Kairos Team, Xiaogang Wang, et al. (arXiv Physical AI)”
Source→“Qiming Zhang, Tao Huang, Zuoyi Fu, Zhisheng Zheng, Yunlong Xi, Feng Lv, Xiaoming Wu — Core Kairos team members (listed among 23 total contributors).”
Source→“Video generation models implementing spatial-temporal attention with Full Attention, such as...Hunyuan1.5”
Source→“fundamentally departing from the prevalent yet disjointed practice of post-training or fine-tuning generic open-domain video generators for downstream embodied control, we pioneer a Native Pre-training Paradigm for Physical AI”
Source→“A prominent example is NVIDIA's Cosmos, which leverages generative video foundation models as digital twins and essential infrastructure for physical AI.”
Source→“Meta's JEPA family (e.g., V-JEPA 2, V-JEPA 2.1, and DINO-world) exemplifies this trajectory.”
Source→“frameworks like Dreamer 4 utilize these models as internal simulators where agents can recursively optimize long-horizon behaviors through imagination.”
Source→AI-extracted from podcast / newsletter / paper summaries. May contain errors.