MemBodied: Recurrent Associative Memory for Vision-Language-Action Models

Problem

Vision-Language-Action models currently lack the capability to retain episode-level information beyond the immediate observation, which limits their effectiveness in sequential decision-making tasks. This paper addresses this gap by proposing a novel architecture that integrates episodic memory into these models. The work is presented as a preprint and has not undergone peer review.

Method

The proposed architecture, MemBodied, incorporates a fixed-size episodic memory designed to enhance the model's ability to recall past interactions and contextual information. Key components of MemBodied include:

• Associative State: This component records interactions across policy calls, allowing the model to maintain a history of actions and observations.
• Episode Anchor: This serves to preserve a compact representation of the initial scene, ensuring that critical context is retained throughout the episode.

The action generation process is conditioned on both the current input and the memory components, enabling the model to leverage past experiences effectively. The evaluation of MemBodied is conducted on five tasks from the RMBench benchmark and the LIBERO-Long suite, although specific details regarding the training compute used are not disclosed.

Results

MemBodied demonstrates significant improvements over various baselines:

• It achieves a mean success rate that is $7.81 imes$ higher than a stateless policy.
• Compared to a vanilla recurrent memory, it shows a mean success rate that is $2.98 imes$ greater.
• The architecture outperforms the strongest memory-augmented baseline by a factor of $1.3 imes$ while utilizing $10 imes$ fewer parameters.
• In the LIBERO-Long suite, MemBodied achieves a success rate of 90.6%, which is a 5.4% improvement over the stateless $π_0$ policy.

Limitations

The authors do not report any limitations in the study, and no obvious limitations are identified in the available text.

Why it matters

The introduction of MemBodied has significant implications for the development of more capable Vision-Language-Action models. By effectively integrating episodic memory, this architecture can enhance the performance of models in tasks requiring long-term context retention, paving the way for advancements in areas such as robotics, interactive AI, and complex decision-making systems.