Adaptive Motion Planning for Multi-fingered Functional Grasp via Force Feedback

TL;DR

Dexterous manipulation of objects with multi-fingered hands under pose uncertainty is studied through an adaptive motion planning framework that uses joint-torque force feedback as part of a reinforcement-learning state. The method employs Soft Actor-Critic (SAC) to learn policies that transition from a pre-grasp to a goal functional grasp without relying on visual feedback, using a force-aware reward structure and a locally linear joint trajectory. Key contributions include a formal problem formulation for functional grasp planning with torque-based sensing, a detailed state-action-reward design, and extensive simulated demonstrations on the ADROIT hand showing improved trajectory smoothness, robustness to torque noise, and adaptation to object location and weight. The work demonstrates human-like flexibility and precision in force-guided manipulation and suggests future multimodal sensing to further enhance dexterity in challenging grasping tasks.

Abstract

Enabling multi-fingered robots to grasp and manipulate objects with human-like dexterity is especially challenging during the dynamic, continuous hand-object interactions. Closed-loop feedback control is essential for dexterous hands to dynamically finetune hand poses when performing precise functional grasps. This work proposes an adaptive motion planning method based on deep reinforcement learning to adjust grasping poses according to real-time feedback from joint torques from pre-grasp to goal grasp. We find the multi-joint torques of the dexterous hand can sense object positions through contacts and collisions, enabling real-time adjustment of grasps to generate varying grasping trajectories for objects in different positions. In our experiments, the performance gap with and without force feedback reveals the important role of force feedback in adaptive manipulation. Our approach utilizing force feedback preliminarily exhibits human-like flexibility, adaptability, and precision.

Figures (8)

• Figure 1: In the scenario where there is uncertainty in the initial position of an object and the pose of the object is unknown during the grasping process, an agent uses hand perception to obtain information about the hand-object interaction state, and makes action decisions accordingly. The rewards obtained from the dexterous hand executing actions will guide the update of the critic-net and further update the actor-net.
• Figure 2: Terminal conditions: The hand moves in the incorrect orientation (Left). The object's displacement surpasses the predetermined limit (Right).
• Figure 3: An example of the contribution of force feedback. Left: Pre-grasp. Middle: Grasping trajectories with force feedback. Right: Grasping trajectories without force feedback.
• Figure 4: Quantitative analysis of the contribution of force feedback.
• Figure 5: Guided by real-time force feedback, diverse paths are continually adjusted, aiming to achieve precise grasps while minimizing object displacement.