How Users Understand Robot Foundation Model Performance through Task Success Rates and Beyond

Technical Summary: How Users Understand Robot Foundation Model Performance through Task Success Rates and Beyond

Problem Statement
Robot Foundation Models (RFMs) promise general-purpose home robotics capable of performing arbitrary tasks. However, a critical gap exists in how non-expert users interpret performance metrics when requesting novel tasks—tasks the RFM was not explicitly trained or evaluated on. While Task Success Rate (TSR) is the standard metric for evaluating RFMs, it is unclear if novices interpret this probabilistic data as intended. Furthermore, the high cost of failure in physical environments necessitates that users understand not just if a robot will succeed, but how it might fail. Current evaluations often lack transparency regarding failure modes, potentially leading to unexpected failures and eroded trust. This paper investigates how non-roboticists interpret RFM performance information, specifically focusing on TSR, failure case descriptions, and the interplay between real historical data and internal performance estimates.

Methodology
The authors conducted a two-part study to evaluate user interpretation of RFM performance data:

1. Online Study (n=112): Participants were presented with real evaluation data from three published RFM research projects (OpenVLA, Baku, and Multimodal Diffusion Transformer) involving 16 distinct manipulation tasks. The study employed a within-subjects design where users evaluated robot capabilities based on four information types presented in various permutations:

   • Estimated Task Success Rate (ETSR): The robot's internal estimate of success probability for the requested task.
   • Estimated Failure Case (EFC): A natural language description of the most likely failure mode for the requested task.
   • Related Task - Task Success Rate (RT-TSR): Historical TSR data from a task deemed "similar" to the user's request.
   • Related Task - Failure Case (RT-FC): Historical failure descriptions from similar tasks.

   Participants rated their trust, comfort, and information sufficiency before and after viewing task videos (which were sampled to reflect actual success/failure rates).

2. In-Person Study (n=14): To examine the effect of embodiment, participants observed a Kinova Gen3 Lite robot autonomously attempting a soup-can placement task in a university lobby. They were provided with the same four information types and asked to evaluate the robot's utility and reliability in a home setting.

Key Contributions
The paper makes three primary contributions:

1. Validation of TSR Interpretation: The study empirically validates that non-expert users interpret TSR in a manner consistent with expert expectations. There is a robust positive correlation between reported TSR values and user-reported trust and confidence in the system.
2. Value of Failure Information: The authors demonstrate that natural language descriptions of failure cases are highly valued by non-experts. These descriptions help users understand risks and prepare for potential failures, suggesting a need to standardize failure case reporting in RFM evaluations.
3. Dual Need for Real Data and Estimates: The research highlights that users desire both real historical data from similar tasks (grounded in reality) and the robot's own estimates for novel tasks. This suggests RFMs must be deployed with access to large histories of evaluations and possess the capability to accurately estimate performance on untested tasks.

Results

• Information Sufficiency: The presence of any information type significantly increased users' perceived sufficiency of information. However, success rates (TSR) had a larger effect size on perceived sufficiency than failure cases.
• Comfort and Trust: User comfort and trust were strongly predicted by high success rates (both ETSR and RT-TSR). Conversely, the presence of failure case descriptions (EFC and RT-FC) did not significantly increase reported comfort, though users still rated them as useful.
• Utility Rankings: A majority of users found all information types useful. Quantitative analysis showed ETSR was reported as more useful than EFC, and RT-TSR was more useful than RT-FC. However, ETSR was not significantly more useful than all other types combined.
• Prediction Accuracy: Only ETSR significantly improved users' ability to correctly predict binary success/failure outcomes (61.9% accuracy), indicating it is the most informative metric for immediate outcome prediction.
• Qualitative Insights: Users employed diverse strategies, such as using thresholds (e.g., rejecting tasks with ETSR < 70%) or triangulating estimates with real data. Many expressed skepticism toward estimates without corroborating real-world evidence.
• Embodiment Effects: The in-person study largely corroborated online findings. However, embodied interactions elicited more specific concerns regarding physical safety (e.g., damage to floors or objects) and practical constraints like speed and workspace requirements.

Significance and Claims
The paper claims that while TSR is a necessary metric, it is insufficient on its own for safe and effective deployment of RFMs in home environments. The authors argue for a shift toward user-centered, interpretable evaluations that include:

• Standardized reporting of failure cases (both common and worst-case scenarios).
• Provision of both real historical data and reliable performance estimates for novel tasks.
• Mechanisms for RFMs to behave predictably or forecast behavior on new tasks.

The authors emphasize that informed users are better equipped to collaborate with, teach, and appropriately trust robots. By bridging the gap between expert evaluation metrics and user understanding, this work aims to inform future RFM evaluation standards and deployment strategies, ensuring that users can make informed decisions about when to supervise, teach, or avoid specific robot tasks. The paper modestly notes that while these findings are derived from specific task domains (kitchen chores, manipulation), the principles of information transparency likely generalize across different robot morphologies and task types.