Workspace Optimization Techniques to Improve Prediction of Human Motion During Human-Robot Collaboration

TL;DR

This work tackles the challenge of predicting human goals in human–robot collaboration under motion uncertainty by actively shaping the shared workspace. It introduces Legible Workspace Generation, which combines physical object rearrangement with AR-based virtual obstacles and optimizes configurations using MAP-Elites to maximize legibility, defined through probabilistic goal inference. Across two human-subject studies in 2D navigation and real tabletop manipulation, the approach improves human-goal prediction accuracy and reduces data requirements for learning, demonstrating higher prediction reliability with less training data. The findings highlight the practical potential of environment design to enable safer, more fluent human–robot teaming in diverse shared-work scenarios.

Abstract

Understanding human intentions is critical for safe and effective human-robot collaboration. While state of the art methods for human goal prediction utilize learned models to account for the uncertainty of human motion data, that data is inherently stochastic and high variance, hindering those models' utility for interactions requiring coordination, including safety-critical or close-proximity tasks. Our key insight is that robot teammates can deliberately configure shared workspaces prior to interaction in order to reduce the variance in human motion, realizing classifier-agnostic improvements in goal prediction. In this work, we present an algorithmic approach for a robot to arrange physical objects and project "virtual obstacles" using augmented reality in shared human-robot workspaces, optimizing for human legibility over a given set of tasks. We compare our approach against other workspace arrangement strategies using two human-subjects studies, one in a virtual 2D navigation domain and the other in a live tabletop manipulation domain involving a robotic manipulator arm. We evaluate the accuracy of human motion prediction models learned from each condition, demonstrating that our workspace optimization technique with virtual obstacles leads to higher robot prediction accuracy using less training data.

Figures (11)

• Figure 1: Workspace configuration affects the robot's ability to correctly predict the human's goal -- the blue square cube. Left: The legible path (dotted) requires the human to take a circuitous route while the natural path (solid) is not legible. Right: Our approach generates a workspace configuration by arranging physical objects and projecting "virtual obstacles" in AR (cyan and red barriers), in order to induce naturally legible paths from the human.
• Figure 2: Our approach for generating workspace configurations that enable accurate human goal predictions. (1) In the initialization phase, we sample random environment layouts to populate the behavior performance map, which stores diverse and high performing solutions. This is followed by the improvement phase where we sample directly from the map and add perturbations to test whether the legibility is improved. (2) In both phases, we compute the legibility of the sampled layout by computing the probability of predicting the correct goal at each stage of the task execution. (3) We compute the features of the sampled layout to determine its location in the map. (4) The map is updated if the legibility score of the sampled layout is better than the existing one.
• Figure 3: (a-d) Overcooked layouts and (e-h) initial cube configurations used in our experiments. The environments generated by our approach, (d) Legible and (h) Both Optimized, optimize object and virtual obstacle (shown in cyan with red edges) placements to elicit legible human motion. (i) The tabletop experiment involves the human and the robot collaboratively placing cubes into the desired configuration -- two columns on the right with a given ordering.
• Figure 4: Behavior performance map generated in the Overcooked domain. MAP-Elites is able to explore legible workspace configurations across combinations of the features: number of obstacles and the ordering of ingredient placements, which improves the best solution found in complex search spaces.
• Figure 5: Overcooked experiment results: (a) Mean accuracy of Bayesian predictor using cost function learned from MaxEnt IRL. Compared to the shortest distance heuristic shown in (b), MaxEnt IRL improved accuracy the most when less than $60\%$ of the trajectory has been observed. The width of the shaded area is the Tukey's Q critical value, so that overlaps indicate no statistical significance. The Legible condition elicits significantly higher accuracy for most cases.