import contextlib import io import json import tempfile import unittest from pathlib import Path from unittest.mock import patch import torch from ml.eval.eval_cl import json_finite, run from ml.eval.diag_collision import replay_first_hits from ml.sim.closed_loop import TeacherOracle, rollout class EvalMetricsTest(unittest.TestCase): def setUp(self): self.tasks = [ {'q0': [0] * 6, 'T': [0] * 3, 'exec_s': 1, 'travel': 5}, {'q0': [0] * 6, 'T': [0] * 3, 'exec_s': None, 'travel': None}, {'q0': [0] * 6, 'T': [0] * 3, 'exec_s': 2, 'travel': 6}, ] self.per = {k: torch.tensor(v) for k, v in { 'ok': [True, False, True], 'done': [True, False, True], 't_done': [0.5, float('nan'), 2.0], 'col': [1.0, 0.0, 0.0], 'travel': [6.0, 0.0, 8.0], 'brakes': [2.0, 4.0, 6.0], 'start_fail': [False, True, False], 'goal_ratio_2s': [0.3, 0.9, 0.1], 'err': [0.2, 200.0, 0.8], 'dls_it': [4, -1, 20], }.items()} def evaluate(self, tasks, bs=2, **kw): offset = 0 def fake_rollout(model, kin, batch, device, **kwargs): nonlocal offset p = {k: v[offset:offset + len(batch)] for k, v in self.per.items()} offset += len(batch) return {'time_budget': 'mixed'}, p with patch('ml.eval.eval_cl.rollout', side_effect=fake_rollout), contextlib.redirect_stdout(io.StringIO()): return run(None, None, tasks, 'cpu', bs=bs, **kw)[0] def test_global_denominators_and_collision_free_arrival(self): result = self.evaluate(self.tasks) self.assertEqual(result['n_ref'], 2) self.assertAlmostEqual(result['success'], 2 / 3, places=6) self.assertAlmostEqual(result['collision_free_success'], 1 / 3, places=6) self.assertAlmostEqual(result['collision_rate'], 1 / 3, places=6) self.assertEqual(result['time_ok_1p3'], 1.0) self.assertEqual(result['time_mean_s'], 1.25) self.assertAlmostEqual(result['travel_ratio_median'], 1.2, places=6) self.assertAlmostEqual(result['goal_ratio_2s'], 0.3, places=6) self.assertEqual(result['err_fail_med'], 200.0) self.assertEqual(result['brakes_mean'], 4.0) def test_empty_conditional_metrics_and_explicit_limit(self): task = {**self.tasks[0], 'exec_s': None, 'travel': None} result = self.evaluate([task], max_sec=0.04) self.assertIsNone(result['err_fail_med']) self.assertIsNone(result['time_ok_1p3']) self.assertIsNone(result['travel_ratio_median']) self.assertEqual(result['time_budget'], 'fixed 0.040 s') self.assertEqual(result['collision_free_success'], 0.0) def test_json_replaces_nonfinite_without_changing_zero(self): value = {'items': [float('nan'), float('inf'), -float('inf'), 0.0]} text = json.dumps(json_finite(value), allow_nan=False) self.assertEqual(json.loads(text), {'items': [None, None, None, 0.0]}) def test_first_hit_probe_detects_collision(self): class PlaneKin: def __init__(self, **kwargs): self.cont = torch.zeros(6, dtype=torch.bool) self.lo, self.hi = torch.full((6,), -10.0), torch.full((6,), 10.0) def gaps(self, q, cap): return {'gm': q[:, :1].expand(-1, 49).clamp(max=cap)} traj = torch.zeros(1, 3, 6) traj[0, :, 0] = torch.tensor([1.0, 1.0, -1.0]) per = {'col': torch.tensor([1.0]), 'first_hit': torch.tensor([0.04]), 'dls_it': torch.tensor([-1]), 'traj': traj} with tempfile.TemporaryDirectory() as folder: trace = Path(folder) / 'trace.pt' torch.save({'tasks': [self.tasks[0]], 'batches': [({}, per, 1)]}, trace) with patch('ml.eval.diag_collision.Kin', PlaneKin), contextlib.redirect_stdout(io.StringIO()): result = replay_first_hits(trace) self.assertEqual(result['reported_collision_tasks'], 1) self.assertEqual(result['replayed_first_hits32'], 1) self.assertEqual(result['replayed_first_hits64'], 1) self.assertEqual(result['rows'][0]['gap64_mm'], [0.5, 0.5, 0.0, -1.0]) self.assertEqual(result['rows'][0]['dense_gap64_min_mm'], -1.0) class DlsSafetyTest(unittest.TestCase): def test_rejected_refinement_retains_safe_candidate(self): class ObstacleKin: def clamp_box(self, q): return q def gaps(self, q, cap): return {'gm': torch.where(q[:, :1] > 0.3, -torch.ones_like(q[:, :1]), torch.ones_like(q[:, :1]))} task = {'q0': [0] * 6, 'T': [2, 0, 0]} traj = torch.zeros(1, 2, 6) traj[0, 1, 0] = 0.1 ref = torch.zeros(1, 1, 6) ref[0, 0, 0] = 1.0 patches = { 'features': lambda *args: torch.zeros(1, 83), 'goal_err': lambda kin, q, T, R: (torch.full((q.shape[0],), 2.0), torch.zeros(q.shape[0])), 'dls_steps': lambda *args, **kwargs: ref, } results = [] with patch.multiple('ml.sim.closed_loop', **patches): for gate in (0, 1): _, per = rollout(None, ObstacleKin(), [task], 'cpu', K=1, safety=True, max_sec=0.02, return_traj=True, use_sub=False, select='first', oracle=TeacherOracle(traj), replan='empty', commit_n=1, safety_partial=1, dls_safety=gate) results.append(per) self.assertEqual(results[0]['col'].item(), 1.0) self.assertEqual(results[0]['dls_it'].item(), 0) self.assertEqual(results[1]['col'].item(), 0.0) self.assertEqual(results[1]['dls_it'].item(), -1) self.assertAlmostEqual(results[1]['traj'][0, 1, 0].item(), 0.1) def test_feedback_is_reduced_only_after_full_rejection(self): class ObstacleKin: def clamp_box(self, q): return q def gaps(self, q, cap): return {'gm': torch.where(q[:, :1] > 0.3, -torch.ones_like(q[:, :1]), torch.ones_like(q[:, :1]))} task = {'q0': [0] * 6, 'T': [10, 0, 0]} traj = torch.zeros(1, 17, 6) traj[0, :, 0] = -0.2 * torch.arange(17) ref = torch.zeros(1, 1, 6) ref[0, 0, 0] = 1.0 patches = { 'features': lambda *args: torch.zeros(1, 83), 'goal_err': lambda kin, q, T, R: (torch.full((q.shape[0],), 10.0), torch.zeros(q.shape[0])), 'dls_steps': lambda *args, **kwargs: ref, } results = [] with patch.multiple('ml.sim.closed_loop', **patches): for gain in (1.0, 0.5): _, per = rollout(None, ObstacleKin(), [task], 'cpu', K=1, safety=True, max_sec=0.04, return_traj=True, use_sub=False, select='first', oracle=TeacherOracle(traj), replan='empty', commit_n=1, safety_partial=1, fb_gain=0.6, fb_block_gain=gain) results.append(per) self.assertEqual(results[0]['brakes'].item(), 2.0) self.assertEqual(results[1]['brakes'].item(), 1.0) self.assertEqual(results[1]['traj'][0, 1, 0].item(), 0.0) self.assertAlmostEqual(results[1]['traj'][0, 2, 0].item(), 0.1) self.assertEqual(results[1]['col'].item(), 0.0) if __name__ == '__main__': unittest.main()