Robots That Take Initiative: A Framework for Building and Evaluating Proactive Robots

Maithili Patel; Sonia Chernova

cs.ROSep 24, 2026

Robots That Take Initiative: A Framework for Building and Evaluating Proactive Robots

Authors: Maithili Patel, Sonia Chernova

Organizations: School of Interactive Computing, Georgia Institute of Technology, USA

Abstract

Effective robot assistance beyond narrow roles and repetitive tasks requires robots to be proactive - to decide what needs to be done rather than waiting to be told. While proactivity is increasingly explored, it lacks a unified formulation, and work in the domain is typically evaluated offline against static human models that cannot capture the effect of a robot's actions on the environment and the user's own behavior. We introduce a unified formalism for proactive robot assistance, organize it into three levels, and provide a framework to address the highest level of unprompted proactive assistance. We then show that offline evaluation overstates performance in this setting, and contribute a closed-loop evaluation with a human model that adapts to the robot. Finally, we present a method, GAP, that instantiates our framework, learning from passive observation to anticipate user goals and act. Under closed-loop evaluation, prior state-of-the-art methods collapse, in some cases adding more work than they save, while GAP remains robust and substantially outperforms them.

Figures & tables

Fig. 1 : Variables and processes in the proactive assistance problem. The fully observed environment evolves under joint human $a^{h}_{t}$ and robot $a^{r}_{t}$ actions, while the hidden user state $h_{t}$ evolves in parallel and shapes the active goal $g_{t}$ at each timestep. We introduce three levels of proactivity for the robot policy with increasing unknowns.

Fig. 3 : Our proactivity method, Grounding, Anticipation, and Planning (GAP), in a closed loop evaluation setup. Closing the evaluation loop accounts for effects (purple arrows), in addition to those that open-loop does (grey arrows)

		Closed-Loop
Method	Offline	LLM User	Scripted User
SLaTe-PRO	0.634	0.216	0.264
STREAK	0.677	0.002	0.296

TABLE II : Offline versus closed-loop F1 scores for existing methods, illustrating the misalignment between the evaluation methods.

	LLM User			Scripted User
Method	Net Saved (%) $\uparrow$	Perturbed (%) $\downarrow$	F1 $\uparrow$	Net Saved (%) $\uparrow$	Perturbed (%) $\downarrow$	F1 $\uparrow$
SLaTe-PRO	2.1 $\pm$ 7.5	18.1 $\pm$ 3.1	0.216 $\pm$ 0.034	16.6 $\pm$ 3.0	11.8 $\pm$ 1.4	0.264 $\pm$ 0.043
STREAK	-159.9 $\pm$ 50.3	10.6 $\pm$ 10.8	0.002 $\pm$ 0.001	20.3 $\pm$ 7.3	48.3 $\pm$ 19.0	0.296 $\pm$ 0.105
Plain LLM	-174.5 $\pm$ 26.5	62.1 $\pm$ 4.5	0.076 $\pm$ 0.022	19.5 $\pm$ 4.8	25.2 $\pm$ 0.8	0.310 $\pm$ 0.064
GAP (ours)	17.7 $\pm$ 1.1	11.7 $\pm$ 1.9	0.388 $\pm$ 0.007	30.8 $\pm$ 1.8	10.1 $\pm$ 0.8	0.452 $\pm$ 0.024
Oracle	44.4 $\pm$ 1.6	8.2 $\pm$ 1.1	0.617 $\pm$ 0.015	51.3 $\pm$ 0.0	8.8 $\pm$ 1.9	0.651 $\pm$ 0.006

TABLE III : Closed-loop performance under the LLM and scripted user, with standard deviation intervals between households.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

	Distribution Shift at Test time only			Distribution Shift during Training (main results)
Method	Net Saved (%) $\uparrow$	Perturbed (%) $\downarrow$	F1 $\uparrow$	Net Saved (%) $\uparrow$	Perturbed (%) $\downarrow$	F1 $\uparrow$
SLaTe-PRO	24.2 $\pm$ 5.4	10.8 $\pm$ 3.0	0.345 $\pm$ 0.071	16.6 $\pm$ 3.0	11.8 $\pm$ 1.4	0.264 $\pm$ 0.043
STREAK	0.4 $\pm$ 0.6	66.3 $\pm$ 57.4	0.009 $\pm$ 0.015	20.3 $\pm$ 7.3	48.3 $\pm$ 19.0	0.296 $\pm$ 0.105
GAP (ours)	36.1 $\pm$ 1.8	16.0 $\pm$ 0.7	0.496 $\pm$ 0.022	30.8 $\pm$ 1.8	10.1 $\pm$ 0.8	0.452 $\pm$ 0.024

Appendix

TABLE IV : Closed-loop performance under the scripted human, with a shifted evaluation distribution but not training distribution, compared to iteratively trained versions where distribution shift affects both phases

Panel of Figure 7 — Fig. 4 : Variation in metrics as we vary $c_{thresh}$ .

Explore similar work

PACT: Proactive Asking for Continual Task Assistance in Human-Robot Collaboration
May 23, 2026Chengbo He, Sheng Li, Chenyang Ma +6Human-Robot Collaboration Proactive
University of Science and Technology Beijing · University of Oxford · 3Tsinghua University
Assistance Without Interruption: A Benchmark and LLM-based Framework for Non-Intrusive Human-Robot Assistance
May 2, 2026Yuedi Zhang, Shuanghao Bai, Wanqi Zhou +4Human-Robot Interaction Assistant
Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University, Xi’an, China · School of Electrical Engineering, Xi’an University of Technology, Xi’an, China
ProAct: A Benchmark and Multimodal Framework for Structure-Aware Proactive Response
Feb 3, 2026Xiaomeng Zhu, Fengming Zhu, Weijie Zhou +8Multimodal Agents Action Selection
Department of Computer Science and Engineering, The Hong Kong University of Science and Technology (HKUST), Hong Kong SAR, China · Tencent, Shenzhen, China · Futian Laboratory, Shenzhen, China +1

May 23, 2026, cs.ROJ/K move · Enter open · S save

PACT: Proactive Asking for Continual Task Assistance in Human-Robot Collaboration

Chengbo He, Sheng Li, Chenyang Ma, Bochao Zou+5

University of Science and Technology Beijing · University of Oxford · 3Tsinghua University

Human-Robot Collaboration Proactive Artificial Intelligence Assistants Interaction History

Open paper AskPDF arXiv

Robotic assistants in long-term human-robot collaboration need to assist users under partial observations while leveraging cross-day interaction history. However, human traits and routines are often unknown at the beginning of collaboration, making passive infer-then-act assistance ineffective and inefficient. To address this challenge, we study a cross-day proactive asking setting for continual task assistance and propose PACT (Proactive Asking for Continual Task Assistance), an ask-or-act framework that determines whether clarification should be sought before taking action. PACT leverages current observations together with accumulated interaction history to evaluate contextual sufficiency, enabling the robot to provide more reliable assistance and progressively adapt to the user over time. We implement its primary learned instantiation using reinforcement learning and evaluate alternative instantiations under the same framework. To assess such behavior, we further introduce a clarification utility metric that quantifies the trade-off between assistance accuracy and the frequency of clarification requests. Experiments in multi-day embodied collaboration scenarios demonstrate that, compared with passive inference baselines, PACT consistently improves both assistance accuracy and clarification utility, highlighting the importance of proactive asking in continual human-robot collaboration.

Abstract

Explore similar work

PACT: Proactive Asking for Continual Task Assistance in Human-Robot Collaboration

Assistance Without Interruption: A Benchmark and LLM-based Framework for Non-Intrusive Human-Robot Assistance

ProAct: A Benchmark and Multimodal Framework for Structure-Aware Proactive Response