Research Analyzer
← Back ICRA 2026

Leveraging Embodied Mechanical Intelligence for Learning Decluttering Tasks

Enrico Turco, Valerio Bo, Chiara Castellani, Gionata Salvietti, Monica Malvezzi, Domenico Prattichizzo, Maria Pozzi

PDF

AI summary

Key figure (auto-extracted from paper)
A custom soft-rigid gripper with a scoop design enables robots to learn decluttering tasks using only a single grasp action, drastically cutting training time and eliminating complex push sequences.
Embodied intelligence Soft robotics Reinforcement learning Decluttering Gripper design Robotic grasping

Problem

Most learning-based decluttering methods rely on rigid grippers and complex multi-step policies, overlooking how specialized gripper morphology can simplify robotic learning.

Approach

The authors train a deep reinforcement learning policy to declutter scenes using three grippers, comparing a simplified grasp-only strategy against a traditional push-grasp approach in simulation and real-world tests.

Key results

  • Soft ScoopGripper achieved higher sample efficiency with a grasp-only policy
  • Scoop morphology enabled non-prehensile motions during grasping
  • Grasp-only policy converged faster than push-grasp across all tests
  • Real-world trials confirmed simulation-based learning advantages

Why it matters

Proves that task-specific mechanical design can drastically reduce the complexity and data needs of robotic learning, guiding efficient gripper development for unstructured environments.

Abstract

In this work, we investigate how a state-of-the-art grasp planner based on deep reinforcement learning performs when applied to a soft-rigid gripper in a decluttering task. The gripper, called Soft ScoopGripper, is endowed with a rigid scoop- shaped part that facilitates the interaction with the environment and with objects. We hypothesize that the clever design of such a gripper can facilitate the learning process, reducing the number of required training steps and eliminating the need for learning non-prehensile actions, such as pushing. To validate our hypothesis, we conducted experiments in both simulated and real- world environments, comparing the selected gripper with a rigid parallel-jaw gripper and a four-fingered soft gripper. Results show that the Soft ScoopGripper learns to effectively declutter scenes using a single action (grasping) instead of two (pushing and grasping). This is due to the fact that the scoop-shaped add- on allows to perform non-prehensile motions during the grasp action.

Index terms

Grasping Grippers and Other End-Effectors Modeling Control and Learning for Soft Robots

Related papers