- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper addresses the limited ability of vision-language-action (VLA) models to retain historical information and reuse experience across tasks in long-horizon robotic manipulation. It proposes Remember Smarter (RS), a memory augmentation module that uses spatial and temporal Mamba to compress multi-view visual histories and organizes successful trajectories in hyperbolic space for asynchronous retrieval. Retrieved experiences are converted into prompt tokens to support action generation. Experiments with π0 show that RS improves overall success on LIBERO-Plus from 53.6% to 70.6% and yields gains on real-robot tasks involving counting, occlusion, and complex task composition.
Critical strengths
- The Mamba-based visual history compressor provides an efficient way to equip VLA models with temporal memory, with computational cost scaling linearly with sequence length.
- The motivation for introducing experience memory is compelling: reusing successful experiences from simpler subtasks offers a promising way to support generalization to complex, long-horizon task compositions.
Critical weaknesses
-
The proposed method is relatively straightforward, combining Mamba-based compression of historical tokens with the introduction of “experience” tokens, and its novelty appears limited. The necessity of the Poincaré VAE also remains unclear. Comparisons with simpler encoders and ablations of variational modeling, hyperbolic geometry, and hierarchical organization are needed.
-
The experimental evaluation has substantial limitations. Improvements over a memory-free baseline, without achieving state-of-the-art performance, are insufficient to establish the method’s superiority. Moreover, LIBERO-Plus primarily evaluates robustness rather than memory-dependent reasoning, making it difficult to attribute the gains to the proposed memory mechanisms. Explicitly memory-dependent tasks are needed. The real-robot experiments include only π0 and component ablations; comparisons with existing memory-augmented methods are necessary to establish the approach’s relative advantages.
Questions for authors
- Experience tokens are masked during policy training and enabled only at inference, while the experience branch is trained separately. Does this introduce a training–inference mismatch, and how does the VLA learn to use these tokens without joint optimization?
- Lines 422–425 appear inconsistent with Figure 5: the text attributes results to π0+RS, while the corresponding values are labeled as π0+VMC in the figure. Please clarify the method labels and reported values, and enlarge the figure’s font for readability.