CoDA's co-evolution loop raises out-of-distribution manipulation success well beyond supervised fine-tuning and the strongest same-budget baseline, both on its own task families and on the external LIBERO-Pro benchmark

roboticsvision-language-actiongenerative-simulationbenchmarkliberoout-of-distribution-generalizationarxivcs-ro

CoDA (arXiv:2610.22362, 2026; see claim-coda-shared-memory-couples-data-engine-policy-optimality-preserving-reward-shaping for the architecture) reports two separate evaluations of its data-policy co-evolution loop. On its own four manipulation task families (pick-and-place, articulated, put-inside, stack/insert), out-of-distribution (OOD) success rises from 5.3% under a supervised-fine-tuning baseline trained on the same seed demonstrations to 27.6% under CoDA — reported as 9.5 percentage points above the strongest closed-loop baseline tested (SimWorld Studio, Kang, Ye et al., 2026, reference [10]) at the same environment-generation budget.

Separately, on LIBERO-Pro — an external robustness-testing benchmark built by a different research group (Zhou, Xu et al., arXiv:2510.03827, 2025, reference [52]) — fifteen rounds of CoDA's generation-and-training loop raise the π0.5 policy's (Physical Intelligence) success from 25% to 84% under position perturbations and from 1% to 76% under instruction perturbations. The paper reports this as 7 and 9 percentage points above ASPIRE (Lu, Wu, Kou et al., arXiv:2607.00272, 2026, reference [18]), the strongest external baseline it compares against on that benchmark, with CoDA leading on the Goal and Spatial suites while ASPIRE remains strongest on the Object suite.

Both the LIBERO-Pro and ASPIRE papers are cited sources within CoDA's own text, not independently fetched or verified against their own published numbers this session; the figures above are CoDA's self-reported comparison, not a cross-checked one.