Created: November 17, 2022
Tags: Simulation, Gelsight, Photometric Stereo, PyBullet
Link: https://arxiv.org/pdf/2109.04027.pdf
Status: Reading
Tactile-based feedback insertion policy with RL that generalizes to objects of unknown geometry. The insertion process is modeled as an episodic policy that iterates between insertion attempts followed by pose corrections.
Explore inserting multiple objects into multiple environments with a single policy.
Object insertion is a classic contact-rich manipulation task. The task remains challenging, especially when considering objects of unknown geometry, which significantly limits the ability to understand the contact configuration between the object and the environment.
Since tactile simulation still is not ready for simulating real-world dynamics of insertion and tactile sensing, they conduct the experiments in real-world.
In total, 500 episodes ( 8 hours of robot time) with initial grasp provided and an automatic resetting mechanism based on object-specific alignment features.
Evaluate RL insertion policy in 4 different approaches: