RoboBenchMart: Benchmarking Robots in Retail Environment

Researchers introduced RoboBenchMart, an open-source simulated benchmark for evaluating robotic systems in retail dark-store environments. The study reveals that current state-of-the-art vision-language-action (VLA) models struggle with complex grocery manipulation tasks, indicating limitations in their generalization across diverse domains beyond tabletop scenarios.

RoboBenchMart addresses a critical gap in robotics research by shifting focus from controlled tabletop environments to realistic retail automation scenarios. The benchmark targets mobile manipulators performing tasks in dense, cluttered grocery settings with items at varying heights and depths—conditions that existing benchmarks rarely capture. This work challenges the assumption that generalist VLAs achieving strong performance in household settings automatically transfer to commercial domains, revealing fundamental limitations in current model architectures and training approaches.

The retail automation space represents significant untapped potential for robotics deployment. Dark stores and automated fulfillment centers face severe labor shortages, making this domain economically compelling. However, the complexity of grocery items—their varied geometries, fragility, and spatial arrangements—creates challenges distinct from typical benchmark scenarios. The authors' finding that state-of-the-art models underperform even on common retail tasks highlights the gap between research achievements and real-world applicability.

For the robotics and AI communities, this benchmark provides essential infrastructure for development and standardization. The release of procedural layout generators and trajectory pipelines enables reproducible evaluation and comparative analysis. Developers and companies building retail automation solutions gain a validated testing framework, reducing deployment risk and accelerating iteration cycles. This work also signals that achieving truly general robotic systems requires domain-specific research rather than purely scaling existing architectures.

Looking forward, the robotics field will likely see increased focus on domain-adaptive learning and multi-task training strategies. Future work may incorporate real-world data collection to bridge the sim-to-real gap, particularly for deformable objects common in grocery environments. The benchmark's open-source nature could catalyze collaborative research, attracting investment and talent to practical automation problems with clear economic returns.

• →RoboBenchMart introduces the first major benchmark specifically designed for evaluating robots in retail dark-store environments with dense object clutter.
• →Current state-of-the-art vision-language-action models show significant performance limitations on grocery manipulation tasks, indicating poor cross-domain generalization.
• →The benchmark includes procedural store generators, trajectory pipelines, and baseline models, providing comprehensive infrastructure for robotics research and development.
• →Retail automation represents a high-impact near-term application domain where robotics could address acute labor shortages and operational inefficiencies.
• →The findings suggest that achieving truly general robotic systems requires domain-specific research and training, not merely scaling existing generalist models.