Jeannette Bohg "Robotic Grasping and Manipulation"

Robotic manipulation requires bridging the gap between perception and action to achieve human-like versatility in unforeseen environments. Current research highlights the shift from open-loop, geometry-avoiding motion planning to closed-loop systems that exploit contact and environmental interaction. A central tension exists between high-bias, structured controllers—which offer interpretability and efficiency—and high-variance, end-to-end learning models that require massive data but capture complex behaviors. Roboticist Jeannette Bohg emphasizes that simulation remains brittle for deformable objects, necessitating better abstractions for state and action spaces to improve sim-to-real transfer. Ultimately, the field struggles with generalization across objects, hardware, and tasks, underscoring the need for robust, adaptive systems that can handle uncertainty without relying solely on massive datasets or perfectly calibrated models.