Replicate training across random seeds

Replicate Instruction-Conditioned Motion Control training across independent random seeds to determine the variability and reproducibility of the reported logged-reward improvements.

Background

Each reported reinforcement-learning configuration is represented by one archived training trajectory. The reported increase in logged positive-reward rate may therefore reflect trajectory-specific randomness in policy sampling, code generation, task ordering, or optimization.

Independent seed replication is needed to assess uncertainty across runs and to establish whether the observed changes generalize beyond the archived trajectories.