Sim, Jaeyun (sungkyunkwan university), Song, Eunseop (Sungkyunkwan University(SKKU), Robotics Innovatory), Lee, Seung Yeon (Sungkyunkwan University), Sim, Jae Yoon (Sungkyunkwan university, AIDIN ROBOTICS), Choi, Hyouk Ryeol (Sungkyunkwan University)

RL-Based Adaptive Admittance Control Via Mass–Damping Modulation in Uncertain Contact Environments

Scheduled for presentation during the Work-In-Progress "Interactive Session I: Work-in-Progress Papers" (ThuPS1), Thursday, July 16, 2026, 10:45−12:15, Event Hall

2026 23rd International Conference on Ubiquitous Robots (UR), July 15-18, 2026, Osaka, Japan

This information is tentative and subject to change. Compiled on October 15, 2026

Abstract

Modern robot manipulation increasingly operates in contact-rich environments with uncertain contact stiffness, where low-level controllers must maintain stable and accurate force regulation. This paper proposes an RL-based adaptive admittance control method that jointly tunes the virtual mass Md and damping Dd without explicit stiffness input. The policy observes force error and motion derivatives and outputs continuous admittance parameters at each control step. For continuous-action learning, we adopt the Normalized Advan- tage Function (NAF) architecture with a bounded exponential reward that penalizes force-tracking error, motion roughness, and control effort. To improve early exploration, the replay buffer is initialized with stabilizing transitions generated by an auxiliary one-step MPC in simulation, while the final action is produced by the learned NAF policy. Simulation and real- robot experiments show improved force tracking, especially under low-to-moderate stiffness conditions, and reduced RMSE in continuous-contact hardware tests compared with fixed admittance control and a mass-only RL baseline.