Abstract
This work investigates how the intricate task of a continuous pick & place (P&P) motion may be learned from humans based on demonstrations and corrections. Due to the complexity of the task, these demonstrations are often slow and even slightly flawed, particularly at moments when multiple aspects (i.e., end-effector movement, orientation, and gripper width) have to be demonstrated at once. Rather than training a person to give better demonstrations, non-expert users are provided with the ability to interactively modify the dynamics of their initial demonstration through teleoperated corrective feedback. This in turn allows them to teach motions outside of their own physical capabilities. In the end, the goal is to obtain a faster but reliable execution of the task. The presented framework learns the desired movement dynamics based on the current Cartesian position with Gaussian Processes (GPs), resulting in a reactive, time-invariant policy. Using GPs also allows online interactive corrections and active disturbance rejection through epistemic uncertainty minimization. The experimental evaluation of the framework is carried out on a Franka-Emika Panda. Tests were performed to determine i) the framework's effectiveness in successfully learning how to quickly pick & place an object, ii) ease of policy correction to environmental changes (i.e., different object sizes and mass), and iii) the framework's usability for non-expert users.
Introduction
In industrial automation, the standard "stop-and-go" approach to pick-and-place tasks—where a robot halts completely to grasp an object before moving again—is inefficient compared to the fluid, continuous motions of human workers. However, teaching robots this high-speed fluidity is challenging. Traditional Learning from Demonstration (LfD) often relies on kinesthetic teaching, where a human physically guides the robot arm. This method is inherently limited by the human's ability to move heavy hardware quickly and precisely. Simply accelerating a slow demonstration often fails because high-speed dynamics can cause the robot to drift or topple objects upon contact.
Methodology: Interactive Gaussian Processes
To overcome these limitations, the researchers developed a framework that decouples the dynamics of the motion from its execution speed. The system utilizes Gaussian Processes (GPs) to create a reactive, time-invariant control policy.
Key technical components include:
- Decoupled Dynamics: The framework learns the path shape and the velocity scaling factor ($\gamma$) separately. This allows users to demonstrate the path slowly and then use a gamepad to "dial up" the speed during practice.
- Minimum Uncertainty Dynamical System (MUDS): To handle the "covariate shift" that occurs at high speeds, the system incorporates a stabilization field. This field uses the GP's epistemic uncertainty to push the robot back toward the demonstrated path when it enters unfamiliar states.
- Minimum Uncertainty Inference: For constraints like gripper width and orientation, the system projects the current state onto the most similar point in the training data, ensuring safe behavior even in unfamiliar workspace regions.
Experimental Results
The framework was evaluated using a Franka-Emika Panda robot on tasks involving various object types and weights.
Performance and Speedup
The system successfully reduced execution times from an average of 11 seconds (human demonstration) to under 3 seconds (autonomous execution). A critical finding was the necessity of the MUDS logic; in ablation studies, the robot achieved a 95% success rate with MUDS, compared to only 45% without it, as the robot otherwise drifted off-course during high-speed maneuvers.
Usability and Generalization
A user study with 10 non-experts demonstrated that participants could teach the robot a successful high-speed task in an average of just 5.4 minutes. The learned policies proved robust, generalizing to different object weights (up to 900g) and slight positional changes without requiring new demonstrations.
Conclusion
By leveraging interactive corrections and uncertainty-aware dynamics, this framework enables non-experts to teach robots motions that exceed their own physical demonstration capabilities. Future work will explore improving the teleoperation interface for orientation corrections, incorporating obstacle avoidance, and adding haptic feedback to help users "feel" the robot's uncertainty during the training phase.
Summary updated