Chinese Teams Showcase Spatial Intelligence and Embodied AI Advances at ECCV 2026

In one paper, Spatial-TTT, teams from Tsinghua University, Tencent Hunyuan and Nanyang Technological University propose a test-time training method for streaming visual input. The approach makes a subset of a 2-billion-parameter model's weights update during inference, creating what the authors call fast weights that serve as compact nonlinear memory. The design keeps 25 percent of full attention layers for global semantic reasoning and replaces the rest with TTT layers to compress long video streams. The researchers say the model outperforms GPT-5 on spatial benchmarks while keeping computation and memory growth linear in sequence length.

Another paper, RoboTracer, comes from researchers at Beihang University, Peking University, the Beijing Academy of Artificial Intelligence and the Institute of Automation of the Chinese Academy of Sciences. They choose not to let a vision-language model directly output joint torques or absolute 3D coordinates. Instead, the model predicts a multi-step trajectory as image coordinates plus metric depth, or (u,v,d), a representation that can be projected into the 3D physical space using camera intrinsics. The team said this lets one model drive different robot bodies without retraining, including a UR5 arm and a Unitree G1 humanoid robot. They built TraceSpatial, a dataset with about 4.5 million samples and 30 million question-answer pairs, and introduced a rule-based process reward combined with GRPO reinforcement learning to sharpen metric-space reasoning. In an interview with Leiphone, Zhou Enshen, a doctoral student at Beihang and a core author, said a good representation is key to helping general VLMs learn 3D efficiently.

The GigaAI lab also had a prominent showing at ECCV 2026, with 13 papers accepted, seven of them focused on spatial intelligence. Together with Tsinghua University, the Chinese Academy of Sciences, Zhejiang University and Shanghai Jiao Tong University, GigaAI proposed VLA-R1, which adds explicit chain-of-thought reasoning and GRPO reinforcement learning to a vision-language-action model before it outputs control commands. The same group introduced OmniNWM, a driving world model that generates panoramic RGB video, depth, semantics and 3D occupancy while also producing dense rewards for collision safety and traffic rules, turning the model into a closed-loop testbed for driving policies. Other papers include ReconPhys, which recovers physical attributes such as stiffness and mass from a single monocular video, and MoGe4D, which synthesizes dynamic 4D scenes from one image.

A separate set of contributions came from Qunhe Technology, Nvidia, Intel, Adobe and universities including Zhejiang University and ETH Zurich. In three ECCV papers they proposed infrastructure for physical AI. One is SPEAR, a simulator built on Unreal Engine that uses runtime reflection and shared memory so Python researchers can access more than 14,000 native engine functions, reaching 73 frames per second at 1080p. Another is Syn-GRPO, which lets a vision model generate its own more difficult perceptual training samples during reinforcement learning, counteracting an entropy collapse that standard RL often suffers with fixed datasets. The third is WalkerBench, an interactive navigation benchmark built on real street views from more than 160 cities, including many non-English-speaking cities to reduce geographic prior memorization; it tests an agent's ability to estimate distances and navigate using only onboard first-person visual input.

The cluster of papers shows a growing consensus among Chinese researchers and companies that spatial AI needs memory, metric geometry and closed-loop interaction rather than simply scaling up context windows or treating images as discrete tokens. Many of the studies were released with code and datasets, and the authors say they expect them to serve as infrastructure for further work in robotics and embodied intelligence. The ECCV 2026 program and the accompanying reports from Leiphone described these results as part of a broader effort to push AI beyond offline 2D perception and toward systems that can operate safely in the physical world.