| """Policy behavior in nominal ROM on a goal set: mean throttle and steering chatter (Sim2Sim, descriptive). |
| Usage: python behavior_stats.py A explore -> results/behavior_<set>_<pool>.json""" |
| import json |
| import sys |
|
|
| import numpy as np |
| import torch |
|
|
| import rom_batch |
| from common import RESULTS, Actor, make_goal_set, policy_registry, relative_obs |
| from run_rom import N_PER_RANGE, POOLS |
|
|
| torch.set_num_threads(2) |
| set_name, pool = sys.argv[1:3] |
| reg = policy_registry() |
| goals = make_goal_set(set_name, N_PER_RANGE[set_name]) |
| gxy = np.array([[g['x'], g['y']] for g in goals]) |
| out = {} |
| for p in POOLS[pool]: |
| actor = Actor(reg[p]['path']) |
| o = rom_batch.simulate(actor, gxy, {}, noise_seed=0, keep_traj=True) |
| tr = o['traj'] |
| S = tr.shape[1] |
| acts = np.stack([actor(relative_obs(*(tr[:, k, i].astype(np.float64) for i in range(4)), gxy[:, 0], gxy[:, 1])).numpy() |
| for k in range(S)], 1) |
| steps = np.where(o['success'], np.round(o['time_s'] / 0.1).astype(int), S) |
| valid = np.arange(S)[None, :] < steps[:, None] |
| ds = np.abs(np.diff(acts[:, :, 1], axis=1))[valid[:, 1:]] |
| out[p] = dict(mean_throttle=float(acts[:, :, 0][valid].mean()), mean_abs_steer_change=float(ds.mean()), |
| share_steer_change_gt_0p1=float((ds > 0.1).mean()), share_saturated_steer=float((np.abs(acts[:, :, 1][valid]) > 0.95).mean())) |
| print(p, {k: round(v, 3) for k, v in out[p].items()}, flush=True) |
| (RESULTS / f'behavior_{set_name}_{pool}.json').write_text(json.dumps(out, indent=1) + '\n') |
|
|
