Robocasa-Mem: Benchmarking Environment and Process Memory in Long-Horizon Mobile Manipulation

TL;DR

RoboCasa-Mem introduces a benchmark with nine mobile manipulation tasks requiring agents to use environment and process memory for long-horizon decision-making. It features Behavior-Conditioned Evaluation to assess consistency with past actions and includes 1,800 demonstrations and a hierarchical memory-augmented baseline policy.

Why It Matters

This benchmark enables more realistic evaluation of memory-aware policies in complex, real-world robot tasks, addressing limitations in prior benchmarks that rely on fixed trajectories. It supports the development of robust, adaptive mobile manipulation systems capable of handling dynamic and unobservable task contexts.

Key Findings

• RoboCasa-Mem evaluates both environment memory (spatial, attribute, relational) and process memory (task progress, scheduling) in mobile manipulation.
• Behavior-Conditioned Evaluation dynamically constrains later actions based on prior robot behavior, reducing reliance on fixed targets and enabling fairer assessment of memory consistency.
• The HMAP baseline uses structured memory and subgoal-conditioned vision-language action generation to achieve effective long-horizon task execution.