ROTATE paper results

Evaluated training runs and evaluation teammates behind every figure and table of ROTATE: Regret-driven Open-ended Training for Ad hoc Teamwork. Code, launchers and the figure-by-figure reproducibility guide live in the ROTATE repository; fetch this dataset with

python scripts/download_results.py            # everything, ~2.7 GB
python scripts/download_results.py --no-checkpoints

Layout

results/<task>/<method>/paper/<timestamp>/
    .hydra/config.yaml        resolved training config
    run.log                   training log
    heldout_eval_metrics/     ego agent evaluated against the heldout teammate set (orbax)
    saved_train_run/          final-iteration checkpoints (open-ended methods, populations)
    ego_train_run/            final ego checkpoint (two-stage methods)
results/<task>/best_heldout_returns.json         normalization bounds
results/<task>/rotate/gae_ps_regret_ts_weighted_lambda_sweep/lambda1-X_lambda2-Y/<ts>/
results/simple_sabotage/rotate/{gae_ps_regret_ts_weighted,traj_regret}_lambda_sweep/...
results/overcooked-v1/<layout>/human_proxy_eval/<ts>/
eval_teammates/<task>/...                        heldout evaluation teammates

Tasks: lbf, overcooked-v1/{cramped_room,asymm_advantages,counter_circuit,coord_ring,forced_coord}, lbf-fov-2, lbf-fov-2-grounded, lbf_2ego_2tm, simple_sabotage. Methods: rotate (and paper:treg, paper:treg_lambda objective ablations), rotate_without_pop, rotate_with_mixed_play, ppo_ego_s5, open_ended_minimax, cole, paired, fcp, brdiv, comedi. Core runs hold 6 seeds; ablations, lambda sweeps, partial-observability and n-agent runs hold 3. Checkpoints are pruned to the last open-ended iteration; analyses that need the full population require retraining with local_logger.save_train_out=true.