Does Imitation Learning Preserve Temporal Robustness in Dexterous Manipulation? An Expert-Learner Comparison Across Task Execution Speeds

Does Imitation Learning Preserve Temporal Robustness in Dexterous Manipulation? An Expert-Learner Comparison Across Task Execution Speeds. By Clinton Enwerem and two others, from the University of Maryland. Published on the first of September. Its code repository is worked out: about one hundred and twenty source files, with a readme and a licence.

Dexterous manipulation policies learned by imitation usually get tested for robustness to changes in the scene, the objects, or the instructions. How they hold up when the task is run faster is examined less often. That leaves a question open: how much of the expert's temporal robustness does a learner keep? The claim here is that it keeps noticeably less, and that matching the expert at normal speed says little about it. Equal task success at nominal speed does not imply that expert performance is preserved across execution speeds.

The test is built to be a fair comparison. Expert and learner face the same task conditions, the same draws of initial conditions, and the same speedup factors. The task is called ParcelStow, a contact-rich job in which the robot acquires a parcel, reorients it, and inserts it. The demonstrations cover the speedup range for the manipulation phases that come after the parcel has been acquired. The expert is scripted. The learner is an Action Chunking with Transformers policy, ACT for short, trained on the expert's demonstrations.

At nominal speed the two look identical: both reach one hundred percent task success. Then they come apart, and they do so inside the range the demonstrations already covered. At the maximum demonstrated speed the expert succeeds eighty-four percent of the time. ACT succeeds fifty-three percent of the time.

That gap is not a fluke of one training run. Two ACT policies with different parameter initializations degrade in a similar way. From nominal speed to the maximum demonstrated speed, one drops by thirty-four percentage points and the other by forty-eight. The expert drops by sixteen.

Looking stage by stage shows where it goes wrong. At the top demonstrated speed, ACT fails forty-seven times, and thirty-five of those failures are insertion misalignments. The handoff experiment sharpens this. Under the relative-motion handoff, every acquisition made by ACT keeps hold of the parcel through reorientation and transfer in free space. Yet only sixty-four percent of those go on to complete the whole task, against ninety-five percent when the expert made the acquisition.

One more pattern holds everywhere. Across all the evaluated policies and speeds, there were four hundred and fourteen acquisitions without force closure, and not one of them completed the task.

The boundaries are narrow and clearly drawn. This is one task, ParcelStow, with one scripted expert and one family of learned policy, ACT, in two initializations. The speeds that matter here are inside the demonstrated range, not beyond it, which is part of what makes the result pointed: the learner has seen these speeds and still loses ground. The comparison is about temporal robustness specifically, not the scene or object variation that evaluations more often cover.

A learner that ties its teacher at normal pace can still be the weaker of the two once the clock speeds up. In this setup the weakness concentrates in insertion, and it shows up even when the grasp itself was carried through cleanly. Code, data and evaluation scripts are released with the work.