Perfect imitation at normal speed breaks down when the task speeds up.

Robots that learn by copying humans or scripted programs often handle changes in objects well, but how do they cope when forced to move faster? Researchers tested this using ParcelStow, a task where a robot grabs, turns, and inserts a parcel. They compared a scripted expert against an imitation learning model called ACT, which was trained on demonstrations covering different speeds after the parcel is grabbed. At normal speed, both achieved 100 percent success. However, when pushed to the fastest demonstrated speed, the expert completed 84 percent of runs, while the learner managed only 53 percent. Most of the learner's failures happened when misaligning the parcel during insertion. Additionally, every time a robot grabbed a parcel without a secure grip called force closure, it failed the task. Just because a robot matches an expert at normal speed does not mean it keeps up when moving fast.

Why it matters

Robotics engineers training robots to mimic experts might assume robustness, but this paper shows failure under speed changes. This affects safety and reliability in real-world tasks where pace varies.

Questions answered