Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
The paper introduces RMBench, a novel simulation benchmark consisting of 9 robotic manipulation tasks designed to evaluate the memory capabilities of robotic policies. The benchmark addresses the limitations of existing benchmarks that do not adequately assess memory-dependent manipulation, a crucial aspect for real-world applications. Through extensive experiments using a proposed modular policy (Mem-0) and existing policies, the authors provide insights into the impact of architectural design choices on memory performance in robotic manipulation.
Turns out, robotic manipulation policies struggle with tasks requiring memory, and this benchmark reveals the architectural design choices that actually matter for improving performance.
Robotic manipulation policies have made rapid progress in recent years, yet most existing approaches give limited consideration to memory capabilities. Consequently, they struggle to solve tasks that require reasoning over historical observations and maintaining task-relevant information over time, which are common requirements in real-world manipulation scenarios. Although several memory-aware policies have been proposed, systematic evaluation of memory-dependent manipulation remains underexplored, and the relationship between architectural design choices and memory performance is still not well understood. To address this gap, we introduce RMBench, a simulation benchmark comprising 9 manipulation tasks that span multiple levels of memory complexity, enabling systematic evaluation of policy memory capabilities. We further propose Mem-0, a modular manipulation policy with explicit memory components designed to support controlled ablation studies. Through extensive simulation and real-world experiments, we identify memory-related limitations in existing policies and provide empirical insights into how architectural design choices influence memory performance. The website is available at https://rmbench.github.io/.