EventVLA: Event-Driven Visual Evidence Memory for Long-Horizon Vision-Language-Action Policies
The episode discusses EventVLA, a framework for long-horizon Vision-Language-Action policies that uses sparse visual evidence memory to solve memory bottlenecks. Hosts detail how EventVLA uses a foresight-driven Keyframe Evidence Memory module to proactively schedule sparse memory writes based on predicted future probabilities, leading to improved performance on complex manipulation tasks.