Is Your Robot's Memory Just a Clock? Temporal Shortcuts in Memory-Augmented VLA Policies
In RoboMME counting tasks, injecting brief pauses or wrist fidgets reduces success rates from 47 to 16 out of 50 episodes on one task and from 44 to 10 on another for released memory VLA policies. The failure is traced to a temporal shortcut where elapsed time predicts progress in demonstrations, and removing time encoding does not cure it because null event frames still occupy slots in time-sampled memory. Indexing memory by events rather than time recovers half or more of the lost success rate without retraining the policy, whereas training with injected pauses fails to remove the failure and impairs counting.
展开解读与证据
In RoboMME counting tasks, injecting brief pauses or wrist fidgets reduces success rates from 47 to 16 out of 50 episodes on one task and from 44 to 10 on another for released memory VLA policies. The failure is traced to a temporal shortcut where elapsed time predicts progress in demonstrations, and removing time encoding does not cure it because null event frames still occupy slots in time-sampled memory. Indexing memory by events rather than time recovers half or more of the lost success rate without retraining the policy, whereas training with injected pauses fails to remove the failure and impairs counting.
为什么值得关注
A memory designed to track what happened should advance with events rather than clock ticks to avoid vulnerability to null events in manipulation tasks.
研究问题
Memory-augmented vision-language-action policies that sample frames uniformly by time step fail to distinguish waiting from working, leading to phantom progress where null events like pauses are counted as task completion.
方法原理
The failure is traced to a temporal shortcut where elapsed time predicts progress in demonstrations, and removing time encoding does not cure it because null event frames still occupy slots in time-sampled memory.
结果与证据
In RoboMME counting tasks, injecting brief pauses or wrist fidgets reduces success rates from 47 to 16 out of 50 episodes on one task and from 44 to 10 on another for released memory VLA policies. Indexing memory by events rather than time recovers half or more of the lost success rate without retraining the policy, whereas training with injected pauses fails to remove the failure and impairs counting.
阅读边界
依据本次返回的正文节选整理,未核验未返回的章节、附录或独立复现。
评论 · —
昵称自填。评论会展示给其他读者,请勿填写私人信息。
正在读取评论…