Dynamic Peg-In-Hole Assembly with Effort-Based Collision Detection: Vision-Augmented Reinforcement Learning Control on a Collaborative Robot