Robot Risk Assessment Study

Semantic-Metric Bayesian Risk Fields

Study Instructions

The four videos on the left show robot arm behavior from different policies: Tele-operated, GR00T, Diffusion Policy, and Risk (Ours). The videos are shuffled to avoid bias. Please rank these videos from safest to riskiest.

Important: All trials are successful completions of the task. Please do not rate trajectories based on task success. Some methods may not release the object at the end, but this should not affect your risk assessment.

What does "least risky" mean? A robot behavior is considered less risky when it not only avoids collisions, but also maintains a safe distance from hazardous object combinations in the event of failure. For example, a cup of water should stay away from a laptop even if the robot arm errors, or a knife should maintain distance from fragile objects.

How to fill out the study: Watch all four videos on the left, then use the form on the right to rank them. For each trial, select which position (Upper Left, Upper Right, Lower Left, Lower Right) corresponds to each ranking (1 = Safest to 4 = Riskiest). Important: Each video must be assigned a rating, and no two videos can have the same rating. Use the Prev/Next buttons to navigate between trials.