Ozyegin University
Autonomous Suturing
Project.

Bimanual object handover is a fundamental capability for collaborative robotic manipulation, particularly in precision-demanding surgical subtasks. This paper investigates the application of Context-Based Echo State Networks (CESNs) as a computationally efficient Learning from Demonstration (LfD) mechanism for autonomous bimanual needle handover. We develop a dual-arm simulation environment in NVIDIA Isaac Sim featuring two Franka Panda manipulators equipped with patient-side manipulator (PSM) tools. Demonstration trajectories are generated programmatically utilizing the cuMotion planner across four varying spatial needle initializations. Addressing high-dimensional control, the CESN architecture receives external task contexts and robot joint feedback to directly predict continuous 11-degree-of-freedom (DoF) joint-position commands for closed-loop execution. We evaluate the model’s performance through offline joint trajectory reconstruction and online simulation rollouts. The results indicate that CESNs can accurately reproduce context dependent bimanual trajectories with low computational overhead, achieving millimeter-level precision and offering a lightweight alternative to high-capacity deep learning models for encoding structured manipulation subtasks.

Learning from Demonstration (LfD) enables policy acquisition from expert trajectories; however, adapting those skills to novel goals or constraints without specifying exact trajectories remains difficult. Reinforcement Learning (RL) can extend LfD in such cases, but conventional LfD–RL approaches are computationally costly: they involve high-dimensional parameter optimization and often require retraining the original LfD policy network during RL. We introduce the Dynamic Adaptive Reservoir Computing (DARC) model, a two-stage learning framework that avoids these limitations. In the first stage, a Context-based Echo State Network (CESN) is trained from demonstrations to capture the base LfD behavior. In the second stage, RL modulates the internal reservoir dynamics through low-dimensional context signals, without modifying the learned LfD weights to generate movements that satisfy new goals or constraints. This architecture preserves previously acquired skills while allowing efficient RL-based adaptation. Experiments on reaching with obstacle avoidance show that DARC achieves higher success rates in skill extension and generalization while requiring less computation than a competitive LfD–RL baseline. The proposed framework further maintains its performance advantage under changes in robot morphology and dynamics. In particular, we demonstrate its scalability on a three dimensional pick-and-place task, where a three-dimensional context signal effectively modulates the reservoir dynamics, generating an eight-dimensional joint-and-gripper command. Overall, the results indicate that integrating reservoir-based LfD with reinforcement-driven dynamical modulation provides an efficient and scalable approach for adapting demonstration-learned behaviors to new goals, constraints, and changes in robot morphology and dynamics.