DFL-TORO: A One-Shot Demonstration Framework for Learning Time-Optimal Robotic Manufacturing Tasks

TL;DR

DFL-TORO introduces a one-shot Demonstration Framework that preprocesses kinesthetic demonstrations to yield time-optimal, noise-free, and jerk-regulated trajectories under robot-kinematic constraints. It uses a two-stage optimization (Time Optimization and Trajectory Generation) on B-Spline representations, followed by a Refinement Phase where a human supervisor can slow down execution and extract per-waypoint tolerances in task space. The approach is validated on FR3 and ABB YuMi, demonstrating substantial reductions in execution time and end-effector jerk, and a DMP case study showing improved generalization when LfD is fed optimized demonstrations. The work prospects include automating default tolerance selection, improving operator-assisted refinement to reduce cognitive load, and exploring AR-assisted supervisory loops for safer real-world deployment.

Abstract

This paper presents DFL-TORO, a novel Demonstration Framework for Learning Time-Optimal Robotic tasks via One-shot kinesthetic demonstration. It aims at optimizing the process of Learning from Demonstration (LfD), applied in the manufacturing sector. As the effectiveness of LfD is challenged by the quality and efficiency of human demonstrations, our approach offers a streamlined method to intuitively capture task requirements from human teachers, by reducing the need for multiple demonstrations. Furthermore, we propose an optimization-based smoothing algorithm that ensures time-optimal and jerk-regulated demonstration trajectories, while also adhering to the robot's kinematic constraints. The result is a significant reduction in noise, thereby boosting the robot's operation efficiency. Evaluations using a Franka Emika Research 3 (FR3) robot for a variety of tasks further substantiate the efficacy of our framework, highlighting its potential to transform kinesthetic demonstrations in contemporary manufacturing environments. Moreover, we take our proposed framework into a real manufacturing setting operated by an ABB YuMi robot and showcase its positive impact on LfD outcomes by performing a case study via Dynamic Movement Primitives (DMPs).

Figures (14)

• Figure 1: Advantages of DFL-TORO, Transforming original kinesthetic demonstration trajectories into time-optimal, noise-free, and jerk-regulated demonstrations with the possibility to independently refine the demonstration velocity profile. DFL-TORO acts as an intermediary layer between capturing demonstrations and feeding them into the LfD algorithm.
• Figure 2: Illustrative examples to clarify the proposed refinement concept. The shaded regions represent the tolerance bounds. Red shaded areas indicate zones where the robot is intuitively expected to slow down based on teleoperated brake commands from the human operator. These regions require narrower tolerance bounds compared to other parts of the trajectory, represented by blue-shaded areas.
• Figure 3: DFL-TORO workflow. Red arrows indicate interactive procedures with the human-robot in the loop.
• Figure 4: Illustration of B-Spline fitting approach for each optimization step. The green lines represent the path and waypoints. Each red arc represents one B-Spline.
• Figure 5: Visual diagram of the refinement phase. $C(t)$ and $s_r(t)$ are obtained during the interactive loop, leading to velocity adjustment and tolerance extraction.