| |
| |
|
|
| """Adaptive sampling utilities for run_eval.py. |
| |
| Decides whether to run another batch of episodes for a task, based on the width |
| of the Bayesian credible interval over the success rate. The estimator is the |
| Beta posterior with uniform prior: p ~ Beta(k+1, n-k+1). |
| |
| Defaults follow the TRI LBM sim protocol (Toyota Research Institute, "A Careful |
| Examination of Large Behavior Models for Multitask Dexterous Manipulation", |
| arXiv:2507.05331), which uses 200 rollouts per task in sim. A 95% CI width of |
| ~0.14 is the worst-case (k/n=0.5) width at n=200, so target_width=0.14 with |
| n_max=200 reproduces their effective precision while letting easy tasks stop |
| earlier. Tighten to 0.10 for publication-grade precision; loosen to ~0.27 to |
| match their 50-rollout real-world protocol. |
| """ |
|
|
| from scipy.stats import beta |
|
|
|
|
| def should_continue_sampling( |
| k: int, |
| n: int, |
| target_width: float = 0.14, |
| n_min: int = 10, |
| n_max: int = 200, |
| ) -> bool: |
| """Return True if the eval should run another batch on this task. |
| |
| Stopping rule depends only on uncertainty (CI width), not on the value of |
| the estimate, so the resulting estimator stays unbiased. |
| """ |
| if n >= n_max: |
| return False |
| if n < n_min: |
| return True |
| lo, hi = beta.ppf([0.025, 0.975], k + 1, n - k + 1) |
| return (hi - lo) > target_width |
|
|
|
|
| def count_task_episodes(episode_results: list[dict], env_name: str) -> tuple[int, int]: |
| """Count (successes, total) for one task from the running episode_results list.""" |
| eps = [ep for ep in episode_results if ep.get("env_name") == env_name] |
| n = len(eps) |
| k = sum(1 for ep in eps if ep.get("success")) |
| return k, n |
|
|
