Generalization of memory designs across robotic manipulation tasks

Determine which memory designs for robotic manipulation policies generalize across tasks, identifying which approaches achieve robust performance across diverse long-horizon, history-dependent manipulation scenarios.

Background

Prior memory-based manipulation methods employ varied policy backbones and heterogeneous evaluation protocols, making comparisons difficult and obscuring general trends. Existing benchmarks have also lacked sufficient diversity and difficulty to stress different kinds of memory, further complicating systematic assessment.

RoboMME is introduced to address this gap by providing a unified benchmark spanning temporal, spatial, object, and procedural memory. The open question concerns which specific memory designs (e.g., representations and integration strategies) actually generalize across tasks when evaluated in a standardized setting.

References

The result supports a precise conclusion: grounded steering helps an Agent turn remembered intent into action, while exact stopping remains unresolved.

— 2AM: Grounding Agent-Side Memory as Guidance for Steerable Action Models in Long-Horizon Manipulation  (2609.11308 - Hu et al., 10 Sep 2026) in Conclusion

While demonstrating the importance of memory, these methods rely on different policy backbones and inconsistent evaluation protocols, making it unclear which memory designs generalize across tasks.

— RoboMME: Benchmarking and Understanding Memory for Robotic Generalist Policies  (2603.04639 - Dai et al., 4 Mar 2026) in Introduction

Given the performance of the proprioception-only policy, we conjecture that this robustness may stem from the policy's maintenance of proprioceptive history.

— Assembling Two Parts in One Hand  (2609.10137 - Pei et al., 9 Sep 2026) in Section 4.3, Quantitative Evaluation with Estimation Error

We remark that this is only one possible representation of salient information, and leave exploration of other approaches to future work.

— Workspace Models: Lightweight Robotic Memory via Saliency-Driven Supervision  (2609.20820 - Dashora et al., 17 Sep 2026) in Section 3, subsection “Representing Salient Information via Key-Patch Reconstruction”

An important open question is how to retain the computational efficiency and flexibility of memory-based learning for longer-horizon model-based inference or continuous real-time adaptation.

— Rapid On-Robot Learning for Dynamic Manipulation Skills: Robot Juggling  (2608.26800 - Lee et al., 27 Aug 2026) in Discussion, paragraph beginning “Beyond the skill-level adaptation considered in this work”

Estimating experience quality without exhaustive labels remains open.

— Beyond End-Task Success: How to Audit Visual Experience Retrieval in Robotics  (2609.26567 - Pathak et al., 22 Sep 2026) in Section 5, “The library changes what the score means,” paragraph beginning “The same rule gets different scores in different libraries.”

Both report gains, which leaves open how to choose between them, so we return to the first principles of partially observable Markov decision processes (POMDPs), in which the agent receives only partial or noisy observations of the state before it acts (Figure 1(a)).

— Divide-and-Remember: Recursive Action-Relevant Memory for Long-Horizon VLA Policies  (2610.00982 - Yu et al., 1 Oct 2026) in Appendix A, Section A (opening paragraph)

The storage choices in \S~\ref{sec:5.6} create an unresolved allocation problem: where should reusable experience reside to limit interference, staleness, and maintenance?

— Physical AI: A Survey of Models, Learning, Systems, and Open Challenges  (2610.21171 - Gao et al., 15 Oct 2026) in Section 10.4.5, Improvement Without Forgetting