Zeva: In-Context Causal Learning for Generalizable Embodied Manipulation

Abstract: Generalizable embodied manipulation remains difficult because robots encounter unseen physical conditions at deployment. Zeva enables a frozen policy to learn in context from its own physical interactions: a Causal Interaction Extractor encodes actions and their induced state changes into causal signals stored in dual-timescale memory, then retrieves relevant experience to guide subsequent actions. Simulation and real-world experiments show state-of-the-art performance, continual improvement as interaction experience accumulates, and transfer of acquired experience across tasks without gradient updates.

@article{zeva,
  title={Zeva: In-Context Causal Learning for Generalizable Embodied Manipulation},
  author={Fu Chen and Xin Ding and Bingjia Huang and Xiangyu Li and Mingju Wang and Jiawei He and Kun Li and Wei Sun and Yunxin Liu and Hao Wu and Ting Cao},
  journal={arXiv preprint arXiv:2608.30880},
  eprint={2608.30880},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2608.30880},

  year={2026}
}