State of VLA Research at ICLR 2026 (mbreuss.github.io)

The International Conference on Learning Representations (ICLR) 2026 has highlighted significant developments in Vision-Language-Action (VLA) models, showcasing a growing trend in research and innovation within this domain. As defined, VLAs involve systems that utilize visual and language inputs to produce robotic actions, but definitions vary, particularly regarding the importance of internet-scale pretraining. This pretraining is seen as crucial for enhancing language instruction adherence and overall task generalization, a potential that current models have yet to fully realize, as they struggle with complex, zero-shot scenarios. The ongoing discussion on what constitutes a VLA, as well as efforts to better categorize and benchmark these models, reflects both the excitement and challenges within the field. Emerging research trends reveal an explosive growth in VLA studies, driven by the increasing convergence of expertise from different domains such as vision and robotics. Key areas of focus include the application of discrete diffusion models for parallel action generation, the integration of embodied reasoning to improve model interpretability and task performance, and advancements in action tokenization to bridge the gap between continuous control values and discrete predictions. Furthermore, efficient VLA training techniques are being explored to make this technology more accessible to researchers with limited computational resources. Collectively, these advancements point towards a mature and dynamic VLA research landscape, presenting myriad opportunities as well as challenges for the AI/ML community.