Run a mock robotics benchmark and review evidence with Inspect Robots
Run the built-in CubePick benchmark with a scripted policy, inspect resolved configuration and scores, and render saved robotics evaluation logs for review.
npx skills add agentskillexchange/skills --skill run-mock-robotics-benchmark-review-evidence-inspect-robots
Invoke this workflow when an operator needs a reproducible baseline for the robotics evaluation pipeline before introducing another policy or embodiment. Inputs are an isolated Python environment, the built-in cubepick-reach task, the scripted policy and the cubepick mock embodiment. Outputs are the saved evaluation log, resolved configuration, status, grader metrics and a local HTML report.
What this skill actually does
Install Inspect Robots in a virtual environment and activate it. Run inspect-robots list to verify registered components. Execute inspect-robots run --task cubepick-reach --policy scripted --embodiment cubepick. Inspect the resulting log with inspect-robots inspect LOG.json, then render it with inspect-robots view LOG.json. Capture task and component identity, package versions, recorded configuration, scores, termination and error details in the review handoff. Preserve the original log alongside the report so the result remains traceable.
The scope boundary is the dependency-light mock benchmark and saved-log review. Explicitly select the mock embodiment rather than relying on workstation defaults. This workflow requires no robot, simulator service, model API or hardware plugin. It does not train a policy, certify real-world performance or authorize physical movement. The upstream framework supports interchangeable policies and embodiments, so multiple agent runtimes can invoke this same CLI procedure. A successful scripted mock run validates the evaluation path only; policy comparisons need separately controlled tasks, seeds and configurations. Pin the package version because upstream warns that its API is still evolving.