Learning policies from high-dimensional visual inputs, such as pixels and point clouds, is crucial in various applications. Visual reinforcement learning is a promising approach that directly trains policies from visual observations, although it faces challenges in sample efficiency and computational costs. This study conducts an empirical comparison of State-to-Visual DAgger — a two-stage framework that initially trains a state policy before adopting online imitation to learn a visual policy — and Visual RL across a diverse set of tasks. We evaluate both methods across 16 tasks from three benchmarks, focusing on their asymptotic performance, sample efficiency, and computational costs. Surprisingly, our findings reveal that State-to-Visual DAgger does not universally outperform Visual RL but shows significant advantages in challenging tasks, offering more consistent performance. In contrast, its benefits in sample efficiency are less pronounced, although it often reduces the overall wall-clock time required for training. Based on our findings, we provide recommendations for practitioners and hope that our results contribute valuable perspectives for future research in visual policy learning.

Dynamic in-hand manipulation remains challenging for soft robotic systems, which have demonstrated advantages in safe, compliant interactions but struggle with highspeed dynamic tasks. In this work, we present SWIFT, a system for learning dynamic tasks using a soft and compliant robotic hand. Unlike previous works that rely on simulation, quasistatic actions, and precise object models, SWIFT learns to spin a pen through trial and error using only real-world data and without requiring explicit knowledge of the pen's physical attributes. With self-labeled trials sampled from the real world, SWIFT discovers the set of pen grasping and spinning primitive parameters that enables a soft hand to spin a pen reliably. After 130 sampled actions per object, SWIFT achieves 10/10 success rate across three pens with different weights and weight distributions, demonstrating generalizability and robustness to changes in object properties. The results highlight the potential for soft robotic end-effectors to perform dynamic tasks. We also demonstrate generalization to different shapes and weights, such as a brush and a screwdriver, with 10/10 and 5/10 success rates, respectively. Videos, data, and code are available at https://soft-spin.github.io.