Self-Supervised Keyframe Discovery for Horizon-Invariant Behavior Cloning

Researchers propose Keyframe Mnemonics, a self-supervised method that discovers information-critical observations, or mnemonics, by learning an objective from randomly sampled past observations and using it as a reward for keyframe selection, then trains a behavior-cloning policy conditioned on those keyframes. On synthetic memory domains the mnemonic-conditioned policies reach 100% success rates and generalize to horizons orders of magnitude beyond training, while on a memory-intensive robot manipulation benchmark across 23 tasks the method achieves a 13.9% average absolute success-rate improvement over the strongest baseline and retains 80% success at 20x longer horizons on a real robot. Code and videos are available at https://keyframe-mnemonics.github.io.

arXiv:2610.10857v1 Announce Type: new Abstract: Behavior cloning BC in non-Markovian environments is a challenging problem because policies have to reason over contextual information over long horizons. Existing policy architectures rely on recurrent or attention-based mechanisms to capture long-term dependencies. However, recurrent models suffer from hidden-state collapse and gradient instability under backpropagation through time, while attention-based models are fundamentally limited by context length. To address these issues, we propose Keyframe Mnemonics, a novel self-supervised method that $\textit{discovers}$ a set of information-critical observations $\textit{mnemonics}$ by learning an objective from randomly sampled past observations and using it as a reward for keyframe selection. We then train a BC policy that conditions on the discovered keyframes to model the action distribution. Under certain task-structure assumptions, our formulation provides context retention guarantees over an infinite horizon, while maintaining a small set of decision-relevant keyframes in the policy's working memory. We evaluate our method on synthetic memory domains, where mnemonic-conditioned BC policies achieve $100$% success rates SR and generalize to horizons orders of magnitude beyond training without performance degradation. Additionally, we evaluate on memory-intensive robot manipulation benchmark, achieving a $13.9$% average absolute SR improvement over the strongest baseline across $23$ tasks and retaining $80$% SR at $20\times$ longer horizons on a real robot. Code and videos are available at https://keyframe-mnemonics.github.io.