Original title: Zeva: In-Context Causal Learning for Generalizable Embodied Manipulation
Fu Chen, Xin Ding, Bingjia Huang, Xiangyu Li +7 · August 2026
Robots struggle to handle new physical environments using only their initial training. Researchers built Zeva to help robots learn from real-world practice without changing their main computer brain. When the robot moves, a tool called the Causal Interaction Extractor notes the action and the resulting change in the world. It turns this into a cause-and-effect signal and saves it in a two-part memory. When choosing its next step, the robot pulls helpful memories and feeds them as clues into its frozen control model. In tests, Zeva beat existing leading models. The robot continually improves itself during deployment without updating its core code. Its success rate grows as it gathers experience, and it can apply lessons learned in one task to help with others.
Imagine you're trying to fix a leaky faucet but the tools don't fit perfectly. Each time you adjust the wrench, you notice how the pipe responds. Normally, you'd have to guess each time. Here, the robot remembers how each adjustment affected the pipe and uses that memory to pick better adjustments next time, without needing a plumber to guide it.
This helps robots adapt to unexpected real-world conditions during jobs like assembly or assistance, making them more reliable without needing constant updates from engineers.
The abstract does not explicitly state limitations, but it focuses on manipulation tasks and the ability to learn without gradient updates during deployment.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Zeva: In-Context Causal Learning for Generalizable Embodied Manipulation — Fu Chen, Xin Ding, Bingjia Huang et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.