Abstract:Soft robots can adaptively interact with unstructured environments. Reinforcement learning‐based controllers that leverage continuum mechanics are promising for solving complex interaction tasks. A closed‐loop pose and force control policy for a dexterous soft manipulator enabling dynamic pushing tasks is trained in simulation. The sim‐to‐real transfer of the control policy showcases the potential of soft robotic manipulation via reinforcement learning. Soft robots can adaptively interact with unstructured environments. However, nonlinear soft material properties challenge modeling and control. Learning‐based controllers that leverage efficient mechanical models are promising for solving complex interaction tasks. This article develops a closed‐loop pose/force controller for a dexterous soft manipulator enabling dynamic pushing tasks using deep reinforcement learning. Force tests investigate the mechanical properties of a soft robot module, resulting in orthogonal forces of 9−13 N. Then, the policy is trained in simulation leveraging a dynamic Cosserat rod model of the soft robot. Domain randomization mitigate the sim‐to‐real gap while careful reward engineering induced pose and force control even without explicit force inputs. Despite the approximate simulation, the sim‐to‐real transfer achieved an average reaching distance of 34±14 mm (8.1%L±3.4%L ), an average orientation error of 0.40±0.29 rad (23°±17° ) and applied pushing forces up to 3 N. Such performance is reasonable for the intended assistive tasks of the manipulator. The experiments uncovered that the soft robot interacting with the environment exhibited torsional and counter‐balancing movements. Although not explicitly enforced, they emerged from the mechanical intelligence of the manipulator. The results demonstrate the potential of soft robotic manipulation via reinforcement learning.

Vision-based reinforcement learning control of soft robot manipulators

Open-Loop Motion Control of a Hydraulic Soft Robotic Arm Using Deep Reinforcement Learning

Vision-Based Robotic Object Grasping—A Deep Reinforcement Learning Approach

Learning-Based Control for Soft Robot–Environment Interaction with Force/Position Tracking Capability

A Motion Planning Method for Visual Servoing Using Deep Reinforcement Learning in Autonomous Robotic Assembly

Robotic arm reinforcement learning control method based on autonomous visual perception

Reinforcement Learning Controllers for Soft Robots using Learned Environments

Distance-directed Target Searching for a Deep Visual Servo SMA Driven Soft Robot Using Reinforcement Learning

Visuo-dynamic self-modelling of soft robotic systems

Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning

Position Control of Cable-Driven Robotic Soft Arm Based on Deep Reinforcement Learning

Model-free control framework for multi-limb soft robots

A Deep Reinforcement Learning Solution for the Low Level Motion Control of a Robot Manipulator System

Dexterous Manipulation from Images: Autonomous Real-World RL via Substep Guidance

Model-Based Reinforcement Learning for Closed-Loop Dynamic Control of Soft Robotic Manipulators

Pushing with Soft Robotic Arms via Deep Reinforcement Learning

Vibration and Position Control of a Two-Link Flexible Manipulator Using Reinforcement Learning

Visual Foresight: Model-Based Deep Reinforcement Learning for Vision-Based Robotic Control

Vision-Based Robotic Arm Control Algorithm Using Deep Reinforcement Learning for Autonomous Objects Grasping

Closed‐Loop Soft Robot Control Frameworks with Coordinated Policies Based on Reinforcement Learning and Proprioceptive Self‐Sensing

Learning to Design and Use Tools for Robotic Manipulation