Benchmark RadarRoboMME-Interference
RoboMME-Interference evaluates robot long-context memory under cross-session interference. The benchmark builds on RoboMME, constructing session histories per query episode with relevant demonstration plus controlled unrelated sessions, and measures task success for memory-augmented vision-language-action models.
- Released
- 2026-06-21
- Readiness
- Inspectable
- Primary field
- Robotics & Autonomous Systems
Why it matters
Existing robot memory benchmarks ignore realistic multi-session interference. RoboMME-Interference quantifies how memory decays with unrelated sessions and whether retrieval mechanisms restore robustness, providing practical decision value for long-deployed robots.
Motivation
Robots deployed in realistic settings will accumulate experience across many sessions and tasks over their deployment.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.