Learning Dexterous Manipulation for a Soft Robotic Hand from Human Demonstration

TL;DR

The paper tackles learning dexterous manipulation with a soft, low-cost hand by leveraging object-centric demonstrations combined with reinforcement learning. It introduces a demonstration-selection mechanism to ignore infeasible human trajectories and extends guided policy search to fuse multiple local controllers into a single generalizable policy. A three-phase learning framework (weight assignment, controller optimization, and supervised learning via BADMM GPS) enables robust, generalizable behaviors despite the soft hand's sensing and actuation limits. Empirical results on the RBO Hand 2 demonstrate successful valve turning, abacus bead manipulation, and bottle grasping, illustrating practical impact for affordable, dexterous manipulation with soft robots.

Abstract

Dexterous multi-fingered hands can accomplish fine manipulation behaviors that are infeasible with simple robotic grippers. However, sophisticated multi-fingered hands are often expensive and fragile. Low-cost soft hands offer an appealing alternative to more conventional devices, but present considerable challenges in sensing and actuation, making them difficult to apply to more complex manipulation tasks. In this paper, we describe an approach to learning from demonstration that can be used to train soft robotic hands to perform dexterous manipulation tasks. Our method uses object-centric demonstrations, where a human demonstrates the desired motion of manipulated objects with their own hands, and the robot autonomously learns to imitate these demonstrations using reinforcement learning. We propose a novel algorithm that allows us to blend and select a subset of the most feasible demonstrations to learn to imitate on the hardware, which we use with an extension of the guided policy search framework to use multiple demonstrations to learn generalizable neural network policies. We demonstrate our approach on the RBO Hand 2, with learned motor skills for turning a valve, manipulating an abacus, and grasping.

Figures (6)

• Figure 1: The RBO Hand 2 manipulating an abacus.
• Figure 2: The RBO Hand 2 is an anthropomorphic pneumatically actuated soft hand consisting of seven actuators. Three of them form the palm and thumb. The air chambers can be physically coupled or actuated separately.
• Figure 3: The three manipulation tasks used in our experiments: Turning a valve, pushing beads on an abacus, and grasping a bottle from a table.
• Figure 4: Comparison of different policies for the valve task: the red line indicates the demonstrated rotation of the valve by $\approx 35 \deg$. On average our method learns the most general feedback strategy. The boxes in the box plot for each test position are our method, single demo baseline 1, single demo baseline 2, hand-designed baseline and oracle plotted from left to right. Although the baselines do well in some positions, the only methods which do consistently well across all positions are our method and the oracle.
• Figure 5: Comparison of the distance moved by the various beads in cm using different policies for the abacus task, at 3 different positions, namely Positions 1, 2 and 3 going downwards. The target column in each table indicates the demonstrated movement of the three beads, and the other columns indicate the mean and standard deviation of other methods. On average our method learns the most general feedback strategy besides the oracle