HANGZHOU, China, July 2, 2026 /PRNewswire/ — Manycore Tech (HKEX: 00068), a leading spatial intelligence company and creator of the SpatialVerse, today announced that three research papers have been accepted to the European Conference on Computer Vision (ECCV) 2026 — one of the three most prestigious peer-reviewed conferences in computer vision and AI, alongside CVPR and ICCV. Spanning simulation, data generation, and spatial evaluation, the three papers represent a systematic demonstration of Manycore Tech’s full-stack capabilities across the Physical AI pipeline.

Betting on Infrastructure, Not Just Algorithms

“As AI moves from the digital world into the physical, the industry’s focus is shifting — from whether large models can understand language, to whether agents can understand space and act within the real world,” said Rui Tang, Chief Scientist at Manycore Tech. “If the defining infrastructure of the large model era was compute, the defining infrastructure of the Physical AI era is simulation and data.”

This is the bet Manycore Tech has been making for seven years. The three ECCV acceptances are its most concrete expression to date — each paper targeting a distinct layer of the Physical AI stack: SPEAR, developed in collaboration with multiple tech giants, advances high-fidelity simulation; Syn-GRPO tackles the fundamental scarcity of diverse 3D training data through a self-evolving data framework; and WalkerBench, paired with the Spatial-IDE framework, establishes the field’s first rigorous real-world benchmark for spatial navigation — and closes the loop from evaluation to physical deployment.

A Full Stack for Physical AI

Taken together, the three papers cover what Manycore Tech sees as the critical pipeline for Physical AI: how the world is simulated, how training data is generated, and how spatial intelligence is measured and validated. The results are telling. WalkerBench reveals that the best models today complete only 24.5% of navigation tasks that humans handle at 70% — a gap that points to a deeper architectural limitation in how current models represent physical space. Spatial-IDE has already been validated through zero-shot deployment on a Unitree G1 humanoid robot, achieving kilometer-scale autonomous navigation on real urban streets.

For seven years, Manycore Tech has built the foundational data and simulation layer that embodied AI depends on — from InteriorNet in 2018 to today’s SpatialVerse. Each of these papers advances a critical piece of that infrastructure, not chasing benchmark headlines, but building the substrate.

From Research to the Real World

Research credibility is necessary — but what ultimately matters is whether the infrastructure holds up in production, at scale, in the real world. Manycore Tech’s SpatialVerse has already accumulated one of the world’s largest repositories of interactive 3D spatial data, anchored by InteriorNet — the largest indoor spatial dataset in existence. Through a dual pipeline that combines physical-space reconstruction and generative synthesis, SpatialVerse continuously converts real-world environments into structured, trainable digital assets at the speed and volume that Physical AI demands.

This infrastructure is already in production. Robotics and autonomous systems companies, including AGIBOT, Galbot, iSquare, and Hesai, have built on this foundation. SpatialVerse capabilities have been cited and acknowledged in joint research by Google and Stanford. The company’s open-source spatial models and InteriorGS dataset have both ranked on HuggingFace’s model and data leaderboards.

“Physical AI is entering a phase where competitive advantage will be determined not by algorithmic novelty alone, but by the depth and quality of the underlying data infrastructure,” said Victor Huang, Co-founder and Chairman of Manycore Tech. “We believe this infrastructure must be open, rigorous, and grounded in the physical world. These three papers are not isolated research contributions — they are a statement of what we stand for, and evidence that we are building it.”

As AI steps off the screen and into the real world, a new race is beginning — and a new class of Physical AI infrastructure builders is stepping into the spotlight.

View original content:https://www.prnewswire.com/news-releases/manycore-tech-announces-three-papers-accepted-to-eccv-2026-unveiling-full-stack-infrastructure-for-physical-ai-302816672.html