Qunhe Technology Lands Three Papers at ECCV 2026, Collaborates with NVIDIA on Physics AI Simulation Platform

Qunhe Technology announced three paper acceptances at ECCV 2026, one of the top three global computer vision conferences alongside CVPR and ICCV. The accepted works span spatial perception and reasoning, reinforcement learning data generation, and high-fidelity physical simulation — all critical areas for physical AI.

As artificial intelligence transitions from the digital to the physical world, the industry focus has shifted from 'can large models understand language?' to 'can agents understand space and act in the real world?'. Spatial understanding, simulation training, and continuous learning are becoming core infrastructure for physical AI development.

The three papers systematically cover the full pipeline from perception to learning to action training. Qunhe Chief Scientist Tang Rui emphasized that high-fidelity simulation platforms have become indispensable for physical AI data production and training support systems.

To address limitations in existing simulation tools — including poor programmability, low data transfer efficiency, and a lack of large-scale structured scene assets — Qunhe jointly proposed SPEAR alongside Adobe, NVIDIA, Apple, and Intel. SPEAR integrates NVIDIA's robot training ecosystem, Qunhe's spatial data and structured scene capabilities, and content assets from Apple and Adobe.

SPEAR exposes over 14,000 native Python APIs for highly programmable control, and can simultaneously output depth maps, surface normals, instance segmentation, semantic segmentation, and material IDs. It seamlessly connects to Qunhe's open-source InteriorAgent and InteriorGS datasets, which already carry SimReady physical attributes like dimensions, collision bodies, and joints.

On the data generation front, Qunhe's Syn-GRPO framework introduces a self-evolving system for reinforcement learning that automatically generates novel training images during training. It combines diverse image generation with a diversity reward mechanism to produce progressively harder training materials.

The company also released WalkerBench, the first interactive spatial intelligence benchmark based on real-world street scenes, covering 161 cities across six continents. Results show that even the strongest current AI models achieve only 24.5% completion, revealing a fundamental mismatch between large models' linear text memory and the need for 3D spatial representation.

Qunhe's SpatialVerse platform is accelerating the full 'data-simulation-evaluation' pipeline, with partnerships including ByteDance, Zhiyuan Robotics, Galaxy General Robotics, and Hesai Technology, alongside global collaborations with Google, NVIDIA, Adobe, and Apple.