Benchmarking memory-augmented robotic generalist policies
(Data updated on 10/16/2026)
Success rate (%) · Higher is better
No methods selected. Enable a tag above to show results.
Params: approximate parameter count. B = billion; M = million.
Train: RoboMME training compute, converted to equivalent H100 GPU-hours using GPU count, training time, and relative peak BF16 throughput. Values are approximate and exclude foundation-model pretraining and inference. Hover over a value for details.
H100 GPU-hours = GPU count × training hours × conversion factor. Relative to H100 SXM: A40 ≈ 0.151×, A100 ≈ 0.315×, H200 SXM = 1×. Reported H100 GPU-hours are used directly.
N/R: not reported. N/A: not applicable. ≈: approximate.
A benchmark for evaluating memory-augmented robotic policies across 16 tasks in four suites: Counting, Permanence, Reference, and Imitation.
Models are evaluated on 800 test episodes (50 per task). Multiple runs with different random seeds are recommended.
(1) Download the dataset and RoboMME repository, (2) train and evaluate with your policy, (3) submit a PR following the submission instructions.