An Oct. 1 arXiv preprint found that, across 12 simulated manipulation tasks, robot policies rarely cut the gap between perturbed and unperturbed runs by half after replanning. Among measurements with a resolved direction, those gaps more often grew than contracted.
Many imitation-trained robot controllers use action chunks: they predict a short sequence of movements and execute several before checking the scene again. This reduces how often the controller must decide what to do, but it leaves any early movement error to propagate while the robot follows the rest of the sequence.
Goyal added translation noise to the first planned action at a magnitude based on the policy's action-prediction error on held-out demonstrations. He compared replaying the rest of the action chunk with replanning after every step, measuring the gap between each perturbed trajectory and a run without the added noise. In the main closed-loop test, a deviation counted as stable only when that gap clearly shrank by at least half over 24 steps, and unstable only when it clearly grew by at least twofold. Smaller changes or an unresolved direction stayed unclassified.
Across 12 tasks in the robomimic, MimicGen and RoboCasa benchmark suites, most measurements were unclassified. Among states where error propagation could be resolved, growth was more common than confident contraction. At the paper’s default twofold-change threshold, only 3 to 88 samples per task were confidently labeled stable under closed-loop replanning; most other samples were uncertain or showed amplification.
The result also depended on how long the study tracked each trajectory. Error growth was often strongest soon after the perturbation and then leveled off, so a short measurement window could overstate how quickly deviations grew across the full sequence. Replanning often removed a clear growth label, but rarely showed that the gap was shrinking.
The result is limited to simulated tests; it does not identify a best chunk length, test physical robots or evaluate recovery training. Goyal argues that imitation-trained policies need direct examples of deviations and corrective actions to learn recovery.
The practical next question is whether explicit recovery training improves task performance on physical robots.
Sources
Aryan Goyal, Measuring the Stability Assumption Behind Action Chunking (Oct. 1, 2026 preprint)
Tony Z. Zhao et al., Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (2023)
Image: Figure 1 from Goyal's paper, licensed CC BY 4.0; arXiv license information.