Qunhe Tech and partners release three ECCV 2026 papers building infrastructure for Physical AI

Three ECCV 2026 papers from Manycore Tech Inc. (群核科技) and collaborators at NVIDIA, Intel, Adobe Research, Zhejiang University, Imperial College London and ETH Zurich describe infrastructure for physical AI. One introduces SPEAR, a high-fidelity embodied simulator whose communication mechanism enables high-fidelity and high-speed rendering. Another presents Syn-GRPO, a self-evolving reinforcement learning framework that uses online synthetic data to mitigate entropy collapse in multimodal large model reinforcement learning. The third examines spatial intelligence evaluation.

Original sources (Chinese)