EventVLA: Event-Driven Visual Evidence Memory for Long-Horizon Vision-Language-Action Policies

EventVLA, from the Shanghai AI Laboratory and universities including the University of Science and Technology of China, Shanghai Jiao Tong University, Dalian University of Technology, Tsinghua University, Peking University, and Huawei Technologies, reframes long-horizon robotic manipulation as sparse visual evidence memory rather than dense history buffering. Built on Qwen3-VL-4B-Instruct with an optimized fine-tuning policy, it combines fixed visual anchors with a Keyframe Evidence Memory module that predicts future keyframe probabilities from transformer hidden states and commits only task-critical frames through thresholding, 1D non-maximum suppression, and FIFO-bounded storage. The paper also introduces RoboTwin-MeM, a diagnostic benchmark of 8 strictly non-Markovian simulation tasks with 430 to 1,544 average steps per episode and required intermediate keyframes ranging from 1 to 5, plus 4 real-world bimanual tasks on the ARX ACONE robot. Across 17 simulation tasks and 4 real-world tasks, EventVLA reports an average success-rate gain of +40% over prior memory-augmented VLAs; on RoboTwin-MeM it reaches 75.2% versus 18.0% for visual anchors alone, while on RMBench it achieves 67.8%, establishing state-of-the-art performance for conventional memory-oriented manipulation. The key technical novelty is foresight-driven memory writing: instead of reacting after evidence disappears, the model predicts which future steps will become keyframes and stores those observations before occlusion, enabling robust long-horizon control without a redundant memory bank.