Supplementary Data — Trained Model Checkpoints and Benchmark Results==================================================================== This archive set accompanies an anonymous manuscript submission. It containsthe final trained model weights and the corresponding benchmark evaluationresults for every model reported in the paper. Contents--------1. arnold-final-checkpoints.tar.gz Trained model weights. Unpacks to: final_checkpoints/ ~2.8 GB. One subdirectory per model (see "Models" below), each containing per-seed subdirectories (seed_0, seed_1, ...). A typical seed directory contains: - rl_model__steps.zip Stable-Baselines3 policy checkpoint - rl_model_vecnormalize__steps.pkl VecNormalize statistics - model_config.json / args.json training + model configuration - _config.json per-task environment configs - vocabulary.json sensorimotor vocabulary (where used) - main_bc_ppo_multi_task.py training entry point (snapshot) - PPO_0/ training logs / tensorboard events 2. arnold-benchmark-results.tar.gz Benchmark evaluation results. Unpacks to: benchmarks/final_benchmarks/ ~4 MB. Same model/seed layout as above; each seed directory holds: - seed___results.json per-task benchmark scores - *.log evaluation run logs 3. arnold-benchmark-results-extra.tar.gz Benchmark evaluation results. Unpacks to: benchmarks/final_benchmarks_extra/ ~11 MB. Contains files used by plotting or other analysis in the paper. Model Description arnold Proposed method (multi-task compositional policy) obc On-policy behavior cloning baseline obc_10_tasks OBC trained on the 10-task subset obc_m / obc_s / obc_xs OBC capacity variants (medium / small / extra-small) obc_st OBC single-task specialists obc_task_sv OBC, task-specific sensorimotor vocabulary variant obc_ppo OBC with PPO fine-tuning obc_wo_obs_norm OBC ablation without observation normalization bc Behavior-cloning baseline ppo_t Transformer-based PPO (no sensorimotor vocabulary) ppo_t_sv Transformer-based PPO with sensorimotor vocabulary mt-ppo Multi-task PPO mt-sac Multi-task SAC