MXI reports a task-conditioned combination of π0.5 policies on all 26 Bench2Dex tasks and four channels (5,200 episodes). It combines seven unchanged released Bench2Dex task checkpoints with our selected fine-tuned policies; the README and JSON disclose each task's source, JAX/PyTorch backend, and retrospective selection.
The matching native evaluation logs are prepared in this HF package. Reported SR is 34.230769%, LSCR 52.091346%, efficiency 0.970985382 over 24 tasks with success, yielding Score 48.554806. Maintainer acceptance and the HF log revision have not yet been assigned.
Contributor: 1011.ai. Contact: lizhehan@1011ai.cn.
Please review the task-conditioned combination, the complete native log evidence, and the source-version comparison. The GitHub and HF PR URLs will be cross-linked when the submission is opened.
<!-- MXI submission-package-sha256: 3aea96d67f3519d75e146d00096ded6c6d00424e8a4c8eb8f998cd8f4c633d86 -->
GitHub metadata PR: https://github.com/Bench2Dex/Bench2D…
