A Chinese-language lab guide for running rule baselines and model-driven MuJoCo manipulation, comparing policies under matched tasks and seeds, and exporting decisions and trajectories.
Original by FBddczEvaluationRepository READMESource reviewed
Before you dive in
What you’ll find in the original
Begin with the no-key rule baseline so physics, tasks, and controls can be validated before any model call.
Compare two or three policies with the same task and seed, then align playback by simulation time rather than inference wait time.
Export configuration, decisions, and trajectories, and do not fabricate scores or zero costs when a provider is disconnected or pricing is unknown.
Worth knowing
Cloud-model runs require provider credentials and may cost money; the default demonstration makes zero model calls.