Chinese Teams Present Embodied AI and Data-Efficient Learning Breakthroughs at IJCAI 2026

Regarding the poor performance of mainstream VLA models in unseen scenarios, independent researcher Dr. Li Quanyi found that models like OpenVLA and π₀ achieve over 95% success on standard benchmarks, but their success rates plummet below 21% on the 20 extrapolation tasks of the libero-ood benchmark (OpenVLA 1%, π₀-fast 17%, UniVLA 21%). By extracting and interpolating "text latent variables" from the model's internal representations, the success rate of π₀ on novel tasks increased from 9% to 83% without any retraining, demonstrating that large models already contain composable generalization capabilities.

Chu Wenhui from Texas A&M University proposed RepSAM, which adapts SAM by fine-tuning only 0.63% of its parameters (about 4 million) after scanning layer-wise representation drift. This achieved 97.9% of the performance of full fine-tuning (632 million parameters, 384 GPU hours) in just 4 hours on a single A100. In transparent object segmentation and physical simulation grasping, mIoU rose from 34.7% to 90.1%, with a 94.4% grasp success rate. The method also runs in 63 ms on an embedded Jetson AGX Orin, making it suitable for industrial edge deployment.

For vision-and-language navigation in continuous environments, a team of Chinese scholars proposed FILD-Nav, which extracts instruction landmark features and injects them into waypoint prediction and topological planning. On the VLN-CE benchmark, success rate improved by 2% and SPL by 3%, while Oracle success in unseen environments improved by 7%, enhancing the robot's ability to align language symbols with physical landmarks in unfamiliar spaces.

Professor Dong Erbao's team from the University of Science and Technology of China (USTC) introduced PECHC, a tactile-driven algorithm for the "last-centimeter" blind spot when vision is occluded. Using hybrid corrective imitation learning, cascaded constraint scheduling, and temporal heterogeneous distillation, it achieved a 97.3% autonomous grasping success rate across 150 real objects in the Visual Dexterity dataset, requiring calibration with only one object. Compared to a domain-randomized PPO baseline, performance improved by 42.8%, showing delicate force control when grasping fragile items.

Professor Xiong Hui's team from the Hong Kong University of Science and Technology (Guangzhou) proposed a distributionally robust reinforcement learning algorithm for unmanned surface vehicles under extreme sea conditions. By replacing the backpropagated gradient, it reduced collision rate by 12.28%, improving navigation stability under strong winds and complex currents.

In data-efficient learning, Wang Yaqing, associate professor at the Beijing Institute of Mathematical Sciences and Applications (BIMSA), delivered an Early Career Spotlight talk at IJCAI 2026. She pointed out that Scaling Law faces bottlenecks from exhausting high-quality data and rising computing costs, while tasks like drug discovery and cold-start recommendation naturally lack big data. She introduced the Data-Efficient Agentic Learning (DEAL) framework, addressing the "data hourglass" problem through experience augmentation, structured architecture, and efficient learning paradigms. Her team also proved theoretically that In-Context Learning is data-dependent meta-learning, where the Transformer's self-attention internalizes the explicit adaptation process of traditional meta-learning.