gentle-manipulation

Sim-to-Real Gentle Manipulation of Deformable and Fragile Objects with Stress-Guided Reinforcement Learning

Under Review

Research Problem

Applying sim-to-real RL for Deformable and Fragile Object Manipulation remains largely unexplored. We propose the first visuomotor learning framework for sim-to-real manipulation of 3D deformable and fragile objects.

Motivational Example

Picking up tofu trained with a vanila RL policy often leads to permanent damage.

Our Approach

We learn gentle and safe manipulation policies using object internal stress, obtained in simulation, as a stress-penalized reward.

We bootstrap it with a rigid-to-soft curriculum and human demonstrations to encourage policy convergence and increase training speed.

Evaluation in Simulation

Learning to gently pick up a tofu block with significantly reduced stress.

(1) A vanilla RL policy generates a high stress spike:

(2) Our approach ensures the stress applied to the tofu remains low:

Task 1: Picking up tofu

Our policy can be transferred from simulation to the real world zero-shot.

(1) Vanilla RL baseline: tofu is significantly damaged.

(2) Our method: tofu remains intact.

Task 2: Pushing tofu to a goal

(1) Vanilla RL baseline: tofu is significantly damaged and goal is not reached.

(2) Our method: tofu remains intact and goal is reached.

Variation of object shape

We replace the previous cylindrical tofu by cubic tofu.

(1) Vanilla RL baseline: tofu is significantly damaged.

(2) Our method: tofu remains intact and goal is reached.

Page updated

Google Sites

Report abuse