Coloring Between the Lines: Personalization in the Null Space of Planning Constraints

TL;DR

The paper addresses continual personalization of generalist robots under safety constraints by introducing Coloring Between the Lines (CBTL), which leverages the null space of CSPs to color between constraint lines and tailor behavior to individual users. CBTL couples CSP generation with learnable personalized constraints, learned online via supervised methods or LLM-based prompts, and uses entropy-based active learning to drive data-efficient adaptation without environment resets. The approach demonstrates strong sample efficiency and generalization across simulations, a web-based user study, and a real-robot feeding system, including reusing learned occlusion constraints to anticipate new scenarios. This work offers a unified, practical pathway to continual, flexible, active, and safe robot personalization with broad applicability to real-world human-robot interaction tasks.

Abstract

Generalist robots must personalize in-the-wild to meet the diverse needs and preferences of long-term users. How can we enable flexible personalization without sacrificing safety or competency? This paper proposes Coloring Between the Lines (CBTL), a method for personalization that exploits the null space of constraint satisfaction problems (CSPs) used in robot planning. CBTL begins with a CSP generator that ensures safe and competent behavior, then incrementally personalizes behavior by learning parameterized constraints from online interaction. By quantifying uncertainty and leveraging the compositionality of planning constraints, CBTL achieves sample-efficient adaptation without environment resets. We evaluate CBTL in (1) three diverse simulation environments; (2) a web-based user study; and (3) a real-robot assisted feeding system, finding that CBTL consistently achieves more effective personalization with fewer interactions than baselines. Our results demonstrate that CBTL provides a unified and practical approach for continual, flexible, active, and safe robot personalization. Website: https://emprise.cs.cornell.edu/cbtl/

Figures (8)

• Figure 1: Real robot demonstration of Coloring Between the Lines (CBTL). The robot personalizes to the user over five meals (four in sim). During the fifth meal, CBTL plans to reposition the plate to prevent violating the user's learned occlusion preferences. It then chooses ketchup based on learned dipping preferences. Finally, CBTL generalizes to avoid occlusions with a new drink.
• Figure 2: Overview of Coloring Between the Lines (CBTL), our method for robot personalization.
• Figure 3: CBTL actively reduces uncertainty and generalizes from feedback.
• Figure 4: Main simulation results. Lines are means and shaded regions are standard errors over 10 seeds. Units are deliberately omitted (time is arbitrary in simulation and user satisfaction is environment-specific). In Cooking, Exploit Only and Epsilon-Greedy obtain equal performance; see Appendix \ref{['app:sim-experiment-details']} for discussion. CBTL (ours) consistently personalizes better and faster than baselines.
• Figure 5: Web-based user study. (a) Example meal. (b) The user is asked which set of choices they prefer on a 5-point Likert scale. (c) Users significantly ($p < 0.005$) prefer CBTL after giving feedback about the first meal. (d) CBTL predicts user responses with increasing accuracy over time.