Most robot foundation models predict actions directly from 2D observations and language instructions, without explicitly modeling how those actions will change the 3D world. WSA₁ asks whether a robot can anticipate future spatial states before acting, so that predicted world changes and action generation constrain one another.

WSA₁ introduces 3D-centric World-Spatial-Action modeling, jointly connecting future 2D visual observations, future 3D world states, and continuous actions. The framework learns spatial regularities from large-scale, heterogeneous demonstrations and transfers them across tasks and robot embodiments.

WSA₁ reaches 93.1% success on RoboTwin 2.0 and 98.2% on LIBERO, while improving real-world robot-control performance by 20% on average over representative baselines. Built around WSA₁, our team placed 4th overall—and 1st among university teams—in the CVPR 2026 RoboChallenge.

Human videos and simulated robot data offer scalable sources for VLA pre-training, yet differences in viewpoint, appearance, and embodiment make this experience difficult to transfer directly to real robots. MiVLA begins with the behavioral structure shared by human hands and robot arms, bringing human and robot demonstrations into a unified pre-training framework.

Through human-robot mutual imitation and bidirectional alignment between their action spaces, MiVLA learns transferable behavioral representations across embodiments. The behavioral fidelity of real human video and the task diversity of simulation therefore complement one another, improving generalization to new viewpoints, visual domains, and robot platforms.

A unified VLA model must preserve the semantic priors of a pretrained vision-language model while also learning physical dynamics and continuous control. Pixel-level future prediction can introduce objective interference and substantial inference cost, so InternVLA-A1.5 organizes understanding, foresight, and action as complementary capabilities with clearly separated roles.

Built on a native VLM backbone, the model continues learning visual question answering and subtask prediction while a lightweight unified expert generates continuous actions. A small set of latent queries distills task-relevant future information during training; the video-generation branch is removed at inference, preserving complex-instruction understanding, long-horizon execution, and real-time control.

The long-tailed Softmax and GeLU activations in Transformers are difficult to represent at low bit widths with uniformly spaced or Log2-constrained quantization levels, making clipping and rounding errors hard to balance. CFQuant removes structural constraints on activation quantization intervals and uses Lloyd-Max iterations to learn quantization values and decision boundaries from a small calibration set, without model retraining.

CFQuant further introduces an efficient scale-shift algorithm (ESA) to reduce the distribution shift between calibration and inference, together with lookup-table matrix multiplication (MM-LUT) for efficient low-bit computation. At W4/A4, Qwen3-8B achieves 67.21% average zero-shot accuracy across six tasks, 25.49 percentage points above PrefixQuant; on DeiT-S, deployment experiments deliver about 8× model compression and 3.7× end-to-end speedup.

Wheat-head segmentation is challenged by limited annotations, shifts in acquisition conditions, and accumulated noise in pseudo-labels. Developed for the Global Wheat competition, this work studies how unlabeled data can be used more effectively to build semantic-segmentation models that remain robust in complex field imagery.

The method combines a SegFormer backbone with two-stage hybrid training, extensive augmentation, and an iterative teacher-student self-training loop that progressively refines pseudo-labels and expands effective supervision. Its emphasis on systematic data use and training design also offers a transferable recipe for other vision tasks with scarce annotations.