robot-agent-exploration/rax is a harness that lets coding agents learn robot tasks through online exploration. The October 24, 2026 commit db744f59 is an indexing change that modifies only two files: README and docs/AGENT_HARNESSES.md. The README added to its news section the news that, on the same date, the LIBERO and LIBERO-PRO branches were released in container and native versions.
The newly indexed libero-pro/container and libero-pro/native branches cover LIBERO and LIBERO-PRO. Both benchmarks are environments that run the four LIBERO suites — the Franka Panda manipulator with a parallel gripper attached, on top of robosuite and MuJoCo — on a single GPU. The container side runs the agent in an offline Docker container, and the official evaluation uses either the LIBERO official initial states, or the LIBERO-PRO conditions that apply perturbations (perturbation, slightly changing the initial placement) to those states. The native side runs the agent CLI as-is, and the isolation is handled by the CLI itself. The coding agents that can be used are listed in the README as Codex CLI, Claude Code, DeepSeek Harness, and Kimi Code, and the harness guide document also added the LIBERO-PRO simulation branches to the list of available backends.
That said, this commit changed only two documents. It did not touch the branch code; it only touched the index table, the harness guide, and reference links, and it contains no evaluation figures such as scores or success rates. The results obtained by running with the LIBERO official initial states and the results from the perturbed conditions are different comparison targets, but which condition's results are described cannot be confirmed from this commit. All that can be confirmed is that the exact versions pinned by each branch are recorded in the README. Basic repository information such as the license is not included in this commit's information.