One-Shot Learning from Demonstration of Contact-Rich Robotic Manipulation by Identifying Physical Interactions

Robotic manipulation has evolved from rule-based systems to learning-driven approaches, with imitation learning (e.g., Dynamic Movement Primitives, Gaussian Processes) enabling better generalization. However, most existing methods implicitly model interactions, lacking explicit physical understanding, which limits performance in contact-rich tasks. Multi-demonstration datasets improve robustness but are costly. Simulation offers data efficiency but lacks tactile feedback. Recent efforts incorporate physical interaction cues, but often treat them as implicit features. The challenge remains: how to efficiently learn complex, contact-intensive tasks from minimal data while ensuring interpretability and adaptability.

The core challenge is enabling robots to learn complex manipulation tasks with minimal demonstrations, particularly in contact-rich scenarios. Existing methods struggle with generalization across environmental variations and often lack explicit modeling of physical interactions, leading to brittle performance. How to identify, segment, and utilize physical contact cues—such as when, where, and what contact occurs—and leverage this information for robust, adaptable task reproduction remains an open problem. Addressing this gap is crucial for deploying robots in unstructured, real-world environments where multiple contact events are common and environment conditions vary.

This work introduces a comprehensive framework that explicitly models the three key aspects of physical interactions—when, where, and what—during demonstration. It employs classification algorithms for interaction types, infers transition conditions based on contact cues, and tracks interaction frames to represent task states. The integration of these components with hybrid position-force control and geometric trajectory blending enables the robot to reproduce complex tasks from a single demonstration with high robustness and generalization. Unlike prior approaches, which often treat interactions implicitly or require extensive data, this method emphasizes physical interpretability and online adaptability, marking a significant advance in few-shot robotic learning.

• �� Record demonstration data: poses, twists, wrenches in end-effector frame.
• �� Segment tasks based on threshold crossing of twist-wrench norms, filtering noise.
• �� Identify interaction frames by minimizing and decoupling mechanical power components.
• �� Classify interaction types (Idle, Constrained, Free, Impeded) using RMS thresholds on twist-wrench pairs.
• �� Infer transition conditions—constraint making, breaking, completion—by observing class changes.
• �� Track interaction frames over time, expressing trajectories in these frames.
• �� Apply geometric offsets to adapt to environmental variations, blending trajectories for smooth transitions.
• �� Use hybrid position-force control, switching control modes based on interaction class.
• �� Monitor transition conditions online to switch task segments dynamically, ensuring robustness.

Experiments involved tasks such as opening locked cabinets, bolt screwing, and surface contouring on a real robotic platform. Each task was demonstrated once, then reproduced autonomously. Metrics included success rate, positional error, and robustness to environment shifts. Baseline comparisons with traditional methods showed significant improvements, especially in environments with geometric variations. Hyperparameters like thresholds for interaction classification and control gains were tuned through ablation studies. The results validated the approach’s effectiveness in complex, contact-rich scenarios with minimal data.

The system achieved over 85% success rate across tasks, with positional errors below 3mm even when environment shifted by 20cm. Robustness tests demonstrated minimal performance degradation under environmental variations. Generalization experiments showed a 20% improvement in task success when geometric offsets were applied, confirming the effectiveness of trajectory blending and geometric parameter adaptation. Ablation studies revealed that explicit interaction modeling contributed over 15% performance gain compared to implicit methods, highlighting the importance of physical interpretability.

This approach is suitable for industrial automation, service robots, and collaborative manufacturing, where rapid task learning and environmental adaptability are critical. It enables robots to perform complex contact tasks with minimal demonstrations, reducing setup time and costs. Future integration with perception systems like vision and tactile sensors will further enhance autonomous learning, making robots more versatile and capable of handling unstructured environments and dynamic objects.

The current framework assumes static, rigid environments, limiting applicability in flexible or dynamic settings. Its performance may decline in high-speed or inertial tasks due to quasi-static assumptions. Interaction classification can be sensitive to sensor noise, affecting transition detection accuracy. Computational complexity may hinder real-time deployment in multi-contact, multi-object scenarios. Future work must address environment dynamics, sensor robustness, and scalability to broader manipulation contexts.