Tsinghua AIR Zeva Lifts Embodied Manipulation Success From 26% to 73% Without Retraining

Tsinghua University, in collaboration with Domain Transform, has introduced Zeva, an embodied manipulation method. Zeva employs in-context causal memory, resulting in a success rate increase from 26% to 73% across various benchmarks. This approach allows performance improvement without the traditional data collection and fine-tuning cycle. Zeva utilizes the Cosmos3 backbone and operates on the principle of separating short-term and long-term memory, preventing noisy failures from impacting long-term memory. The process includes in-context policy injection, making site adaptation a contextual engineering challenge rather than traditional training.