cbb/Robot_Arm
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
ee7fe829ac1f832c6e57e7759ff778c992deb964
Robot_Arm/tests/test_eval_cl.py
T

718 lines
38 KiB
Python
Raw Normal View History

exp: 轮次29——全拒后减半反馈到达55.1%，补齐无碰撞指标与首撞回放
2026-10-20 17:06:53 +08:00
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
exp+fix: 轮次31——确定界连续认证抓到段内碰撞，但单段P50 22.9ms进不了20ms；自适应cap收益经交错复测撤回
2026-10-20 19:13:40 +08:00
from ml.eval.diag_collision import replay_certificates, replay_dense, replay_first_hits
exp: 轮次29——全拒后减半反馈到达55.1%，补齐无碰撞指标与首撞回放
2026-10-20 17:06:53 +08:00
from ml.sim.closed_loop import TeacherOracle, rollout
exp+fix: 轮次31——确定界连续认证抓到段内碰撞，但单段P50 22.9ms进不了20ms；自适应cap收益经交错复测撤回
2026-10-20 19:13:40 +08:00
from ml.common.kin_torch import Kin, seg_seg_dist
exp: 轮次29——全拒后减半反馈到达55.1%，补齐无碰撞指标与首撞回放
2026-10-20 17:06:53 +08:00
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],
exp: 轮次33(上)——实际执行速度做成一等指标并自证；默认路径逐位复现第30轮，256题全部超VMAX
2026-10-20 19:48:23 +08:00
'speed_ratio': [1.66, 0.9, 1.2], 'accel_ratio': [0.3, 0.2, 0.5],
exp: 轮次29——全拒后减半反馈到达55.1%，补齐无碰撞指标与首撞回放
2026-10-20 17:06:53 +08:00
}.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)
exp: 轮次33(上)——实际执行速度做成一等指标并自证；默认路径逐位复现第30轮，256题全部超VMAX
2026-10-20 19:48:23 +08:00
# 极值/分位按逐题值跨批汇总，不能进按 n 加权求均值的批聚合
self.assertAlmostEqual(result['speed_ratio_p50'], 1.2, places=6)
self.assertAlmostEqual(result['speed_ratio_p95'], 1.614, places=6)
self.assertAlmostEqual(result['speed_ratio_max'], 1.66, places=5)
self.assertAlmostEqual(result['over_vmax_rate'], 2 / 3, places=6) # 1.66 与 1.2 都算超 VMAX，0.9 不算
self.assertAlmostEqual(result['accel_ratio_max'], 0.5, places=6)
exp: 轮次29——全拒后减半反馈到达55.1%，补齐无碰撞指标与首撞回放
2026-10-20 17:06:53 +08:00
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)
fix+exp: 轮次30——统一认证与执行限位，到达57.4%不变；加密回放仍检出段内碰撞
2026-10-20 17:26:20 +08:00
def test_dense_probe_detects_between_checkpoints_and_respects_deadline(self):
class ThinObstacle:
fp = 'probe'
def __init__(self, **kwargs):
pass
def gaps(self, q, cap):
return {'gm': ((q[:, :1] - 0.25).abs() - 0.05).clamp(max=cap)}
traj = torch.zeros(4, 3, 6)
traj[:, :, 0] = torch.tensor([[0, 1, 1], [0, -0.5, -1], [0, 0, 0.25], [0, 0, 0.25]])
per = {'traj': traj, 'col': torch.zeros(4), 'limit': torch.tensor([0.04, 0.04, 0.02, 0.04]),
't_done': torch.tensor([float('nan')] * 3 + [0.02])}
with tempfile.TemporaryDirectory() as folder:
trace = Path(folder) / 'trace.pt'
output = Path(folder) / 'audit.json'
torch.save({'kin_fp': 'probe', 'tasks': [{}] * 4, 'batches': [({}, per, 4)]}, trace)
with patch('ml.eval.diag_collision.Kin', ThinObstacle), contextlib.redirect_stdout(io.StringIO()):
result = replay_dense(trace, [0, 1, 2, 3], 0.01, out=output)
with self.assertRaisesRegex(ValueError, 'valid task indices'):
replay_dense(trace, [4])
with self.assertRaisesRegex(ValueError, 'positive max_step_deg'):
replay_dense(trace, [0], 0.0)
self.assertEqual(json.loads(output.read_text())['collision_tasks'], 1)
self.assertEqual(result['collision_tasks'], 1)
self.assertEqual(result['rows'][0]['collision_segments'], 1)
self.assertAlmostEqual(result['rows'][0]['min_gap64_mm'], -0.05)
self.assertGreater(result['rows'][1]['min_gap64_mm'], 0)
for row in result['rows'][2:]:
self.assertGreater(row['min_gap64_mm'], 0)
self.assertEqual(row['steps'], 1)
exp+fix: 轮次31——确定界连续认证抓到段内碰撞，但单段P50 22.9ms进不了20ms；自适应cap收益经交错复测撤回
2026-10-20 19:13:40 +08:00
def test_certificate_replay_samples_actual_movement_before_deadlines(self):
kin = SegmentCertificateTest.ThinObstacle()
kin.fp = 'probe'
traj = torch.zeros(3, 4, 6)
traj[:, :, 0] = torch.tensor([[0, 1, 1, 1], [0, -0.5, -1, -1], [0, 0, 1, 1]])
per = {'traj': traj, 'limit': torch.tensor([0.06, 0.02, 0.02]),
't_done': torch.tensor([float('nan'), 0.02, float('nan')]),
'plans': {'it': torch.arange(3), 'commit': torch.zeros(3, 3, 1, 6),
'certified': torch.ones(3, 3, 1, dtype=torch.bool), 'gaps': torch.full((3, 3, 2), 0.5)}}
with tempfile.TemporaryDirectory() as folder:
trace, output = Path(folder) / 'trace.pt', Path(folder) / 'audit.json'
torch.save({'kin_fp': 'probe', 'tasks': [{}] * 3, 'batches': [({}, per, 3)]}, trace)
with patch('ml.eval.diag_collision.Kin', return_value=kin), contextlib.redirect_stdout(io.StringIO()):
result = replay_certificates(trace, [0, 1, 2], out=output, repeats=1, extra_segments=[(0, 0)])
with self.assertRaisesRegex(ValueError, 'extra segment'):
replay_certificates(trace, [1], repeats=1, extra_segments=[(1, 1)])
self.assertEqual(json.loads(output.read_text())['dense_collision_segments'], 1)
self.assertEqual([(r['i'], r['start_it']) for r in result['rows']], [(0, 0), (1, 0)])
self.assertEqual(result['budgets']['32']['safe'], 1)
self.assertEqual(result['budgets']['32']['collision'], 1)
self.assertEqual(result['accepted_dense_collision_segments'], 0)
self.assertTrue(result['batch_single_match'])
self.assertEqual(result['timing']['calls'], 2)
exp: 轮次29——全拒后减半反馈到达55.1%，补齐无碰撞指标与首撞回放
2026-10-20 17:06:53 +08:00
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)
fix+exp: 轮次30——统一认证与执行限位，到达57.4%不变；加密回放仍检出段内碰撞
2026-10-20 17:26:20 +08:00
class ClampedPathSafetyTest(unittest.TestCase):
def rollout_candidate(self, increments, safety_partial, goal_error=10.0, **kwargs):
class ObstacleKin:
def clamp_box(self, q):
clamped = q.clone()
clamped[..., 4] = clamped[..., 4].clamp(max=0.0)
return clamped
def gaps(self, q, cap):
collision = (q[:, :1] > 0.1) & (q[:, 4:5] < 0.2)
return {'gm': torch.where(collision, -torch.ones_like(q[:, :1]), torch.ones_like(q[:, :1]))}
task = {'q0': [0] * 6, 'T': [goal_error, 0, 0]}
commit_n = len(increments)
traj = torch.zeros(1, commit_n + 1, 6)
traj[0, 1:] = torch.tensor(increments).cumsum(0)
patches = {
'features': lambda *args: torch.zeros(1, 83),
'goal_err': lambda kin, q, T, R: (torch.full((q.shape[0],), goal_error), torch.zeros(q.shape[0])),
}
with patch.multiple('ml.sim.closed_loop', **patches):
_, per = rollout(None, ObstacleKin(), [task], 'cpu', K=1, safety=True,
max_sec=0.02 * commit_n, return_traj=True, use_sub=False, select='first',
oracle=TeacherOracle(traj), replan='empty', commit_n=commit_n,
safety_partial=safety_partial, **kwargs)
return per
def test_full_safety_rejects_collision_after_joint_clamp(self):
# Raw endpoint/midpoint are safe, but the J5 upper limit puts the endpoint at (0.2, 0).
per = self.rollout_candidate([[0.2, 0, 0, 0, 0.6, 0]], safety_partial=0)
self.assertEqual(per['col'].item(), 0.0)
torch.testing.assert_close(per['traj'], torch.zeros(1, 2, 6), rtol=0, atol=0)
def test_partial_safety_rejects_collision_after_joint_clamp(self):
per = self.rollout_candidate([[0.2, 0, 0, 0, 0.6, 0]], safety_partial=1)
self.assertEqual(per['col'].item(), 0.0)
torch.testing.assert_close(per['traj'], torch.zeros(1, 2, 6), rtol=0, atol=0)
def test_partial_safety_keeps_only_clamped_safe_prefix(self):
per = self.rollout_candidate([
[0.05, 0, 0, 0, 0.6, 0],
[0.2, 0, 0, 0, 0.6, 0],
], safety_partial=1)
self.assertAlmostEqual(per['traj'][0, -1, 0].item(), 0.05)
self.assertEqual(per['traj'][0, -1, 4].item(), 0.0)
self.assertEqual(per['col'].item(), 0.0)
expected = torch.zeros(1, 3, 6)
expected[0, 1:, 0] = 0.05
torch.testing.assert_close(per['traj'], expected)
torch.testing.assert_close(per['plans']['commit'][0, 0, 1], torch.zeros(6), rtol=0, atol=0)
def test_near_dls_rejects_clamped_collision_and_keeps_candidate(self):
candidate = [[0.05, 0, 0, 0, 0.6, 0]]
ref = torch.tensor([[[0.2, 0, 0, 0, 0.6, 0]]])
for safety_partial in (0, 1):
with self.subTest(safety_partial=safety_partial):
with patch('ml.sim.closed_loop.dls_steps', return_value=ref) as dls:
per = self.rollout_candidate(candidate, safety_partial, goal_error=2.0, dls_safety=1)
dls.assert_called_once()
self.assertEqual(per['col'].item(), 0.0)
self.assertEqual(per['dls_it'].item(), -1)
self.assertEqual(per['brakes'].item(), 0.0)
expected = torch.zeros(1, 2, 6)
expected[0, 1, 0] = 0.05
torch.testing.assert_close(per['traj'], expected)
torch.testing.assert_close(per['plans']['commit'], torch.tensor(candidate)[None, None])
self.assertTrue(per['plans']['certified'].all().item())
def test_r3_clamped_queue_origin_prefix_and_certificates_match_execution(self):
class SafeKin:
def clamp_box(self, q):
clamped = q.clone()
clamped[..., 4] = clamped[..., 4].clamp(max=0.0)
return clamped
def gaps(self, q, cap):
return {'gm': torch.ones_like(q[:, :1])}
prefixes = []
def capture_features(kin, q, qp, T, R, pre, r, sub, sf):
prefixes.append((pre.clone(), int(r.item())))
return torch.zeros(1, 83)
teacher = torch.zeros(1, 25, 6)
teacher[0, :, 0] = 0.02 * torch.arange(25)
teacher[0, :, 4] = 0.6 * torch.arange(25)
task = {'q0': [0] * 6, 'T': [10, 0, 0]}
patches = {
'features': capture_features,
'goal_err': lambda kin, q, T, R: (torch.full((q.shape[0],), 10.0), torch.zeros(q.shape[0])),
}
with patch.multiple('ml.sim.closed_loop', **patches):
_, per = rollout(None, SafeKin(), [task], 'cpu', K=1, safety=True,
max_sec=0.16, return_traj=True, use_sub=False, select='first',
oracle=TeacherOracle(teacher), replan='r3', commit_n=3, safety_partial=1)
self.assertEqual([r for _, r in prefixes], [0, 2, 3, 3])
for pre, r in prefixes[1:]:
torch.testing.assert_close(pre[..., 4], torch.zeros(1, 3), rtol=0, atol=0)
torch.testing.assert_close(pre[:, :r, 0], torch.full((1, r), 0.02))
self.assertEqual(per['col'].item(), 0.0)
expected = torch.zeros(1, 9, 6)
expected[0, :, 0] = 0.02 * torch.arange(9)
torch.testing.assert_close(per['traj'], expected)
plans = per['plans']
self.assertEqual(plans['it'].tolist(), [0, 3, 6, 9])
torch.testing.assert_close(plans['commit'][..., 4], torch.full((1, 4, 3), 0.6))
self.assertTrue(plans['certified'].all().item())
self.assertTrue((plans['gaps'] > 0).all().item())
checked = 0
for i, start in enumerate(plans['it'].tolist()):
n = min(3, 8 - start)
if n <= 0:
continue
torch.testing.assert_close(plans['qe'][:, i], per['traj'][:, start], rtol=0, atol=0)
before = per['traj'][:, start:start + n]
after = per['traj'][:, start + 1:start + n + 1]
torch.testing.assert_close(plans['points'][:, i, :n], after, rtol=0, atol=0)
torch.testing.assert_close(plans['midpoints'][:, i, :n], before + 0.5 * (after - before),
rtol=0, atol=0)
checked += n
self.assertEqual(checked, 8)
def rollout_fallback(self, safety_partial, midpoint_collision=False):
class ObstacleKin:
def clamp_box(self, q):
return q
def gaps(self, q, cap):
collision = q[:, :1] > 0.3
if midpoint_collision:
collision |= (q[:, :1] > 0.09) & (q[:, :1] < 0.11)
return {'gm': torch.where(collision, -torch.ones_like(q[:, :1]), torch.ones_like(q[:, :1]))}
teacher = torch.zeros(1, 4, 6)
teacher[0, :, 0] = torch.arange(4)
ref = torch.zeros(1, 1, 6)
ref[0, 0, 0] = 0.2
task = {'q0': [0] * 6, 'T': [10, 0, 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,
}
with patch.multiple('ml.sim.closed_loop', **patches):
_, per = rollout(None, ObstacleKin(), [task], 'cpu', K=1, safety=True,
max_sec=0.06, return_traj=True, use_sub=False, select='first',
oracle=TeacherOracle(teacher), replan='empty', commit_n=3,
safety_partial=safety_partial, fb_gain=1.0, fb_fallback=1.0)
return per
def test_fallback_executes_one_safe_dls_step_then_certified_zeros(self):
for safety_partial in (0, 1):
with self.subTest(safety_partial=safety_partial):
per = self.rollout_fallback(safety_partial)
self.assertEqual(per['col'].item(), 0.0)
self.assertEqual(per['brakes'].item(), 1.0)
expected = torch.zeros(1, 4, 6)
expected[0, 1:, 0] = 0.2
torch.testing.assert_close(per['traj'], expected)
commit = torch.zeros(1, 1, 3, 6)
commit[0, 0, 0, 0] = 0.2
plans = per['plans']
self.assertEqual(plans['it'].tolist(), [0])
torch.testing.assert_close(plans['commit'], commit, rtol=0, atol=0)
self.assertTrue(plans['certified'].all().item())
self.assertTrue((plans['gaps'] > 0).all().item())
torch.testing.assert_close(plans['points'][:, 0], expected[:, 1:])
torch.testing.assert_close(plans['midpoints'][:, 0],
expected[:, :-1] + 0.5 * (expected[:, 1:] - expected[:, :-1]))
def test_fallback_rejects_safe_endpoint_with_colliding_midpoint(self):
for safety_partial in (0, 1):
with self.subTest(safety_partial=safety_partial):
per = self.rollout_fallback(safety_partial, midpoint_collision=True)
self.assertEqual(per['col'].item(), 0.0)
self.assertEqual(per['brakes'].item(), 1.0)
torch.testing.assert_close(per['traj'], torch.zeros(1, 4, 6), rtol=0, atol=0)
torch.testing.assert_close(per['plans']['commit'], torch.zeros(1, 1, 3, 6), rtol=0, atol=0)
self.assertFalse(per['plans']['certified'].any().item())
exp+fix: 轮次31——确定界连续认证抓到段内碰撞，但单段P50 22.9ms进不了20ms；自适应cap收益经交错复测撤回
2026-10-20 19:13:40 +08:00
class SegmentCertificateTest(unittest.TestCase):
class ThinObstacle(Kin):
def __init__(self):
self.dtype = torch.float64
self.lo, self.hi = torch.full((6,), -5.0), torch.full((6,), 5.0)
self.cont = torch.zeros(6, dtype=torch.bool)
self.cert_groups = torch.tensor([0])
self.cert_lever = torch.tensor([[1.0, 0, 0, 0, 0, 0]], dtype=self.dtype)
def gaps(self, q, cap):
return {'gm': ((q[:, :1] - 0.25).abs() - 0.05).clamp(max=cap)}
def test_bound_detects_hidden_collision_and_accepts_clear_segment(self):
a, b = torch.zeros(3, 6, dtype=torch.float64), torch.zeros(3, 6, dtype=torch.float64)
a[:, 0] = torch.tensor([0.0, -1.0, 0.25])
b[:, 0] = torch.tensor([1.0, 0.0, 0.25])
result = self.ThinObstacle().certify_segments(a, b)
self.assertEqual(result['safe'].tolist(), [False, True, False])
self.assertEqual(result['collision'].tolist(), [True, False, True])
self.assertEqual(result['queries'].tolist(), [4, 2, 2])
def test_budget_and_depth_exhaustion_do_not_accept_unknown_segment(self):
a, b = torch.zeros(1, 6, dtype=torch.float64), torch.zeros(1, 6, dtype=torch.float64)
b[:, 0] = 1.0
kin = self.ThinObstacle()
for kwargs in ({'max_queries': 2}, {'max_depth': 0}):
with self.subTest(**kwargs):
result = kin.certify_segments(a, b, **kwargs)
self.assertFalse(result['safe'].item())
self.assertTrue(result['unresolved'].item())
self.assertEqual(result['queries'].item(), 2)
exp: 轮次33——限幅实测要付5.1pp到达（57.4→52.3%）；GPU真实周期36.7ms≈27Hz，瓶颈是重复几何而非推理
2026-10-20 20:32:19 +08:00
def test_shape_mismatch_between_endpoints_is_rejected(self):
a = torch.zeros(1, 6, dtype=torch.float64)
b = torch.zeros(2, 6, dtype=torch.float64)
kin = self.ThinObstacle()
for method in (kin.certify_segments, kin.certify_uniform):
with self.subTest(method=method.__name__):
with self.assertRaisesRegex(ValueError, 'identical shape'):
method(a, b)
exp+fix: 轮次31——确定界连续认证抓到段内碰撞，但单段P50 22.9ms进不了20ms；自适应cap收益经交错复测撤回
2026-10-20 19:13:40 +08:00
def test_budget_exhaustion_scales_with_gap_and_bound(self):
class ConstantGap(self.ThinObstacle):
def gaps(self, q, cap):
return {'gm': torch.full((q.shape[0], 1), 0.1, dtype=q.dtype, device=q.device)}
# 界取 1 mm/度：1° 段要证明无碰撞，必须细分到位移界 < 两端间隙之和(0.2 mm)。
kin = ConstantGap()
kin.cert_lever = torch.ones(1, 6, dtype=torch.float64)
a, b = torch.zeros(1, 6, dtype=torch.float64), torch.zeros(1, 6, dtype=torch.float64)
b[:, 0] = 1.0
for budget, safe, queries in ((4, False, 3), (8, False, 5), (9, True, 9)):
with self.subTest(budget=budget):
result = kin.certify_segments(a, b, max_queries=budget)
self.assertEqual(bool(result['safe'].item()), safe)
self.assertEqual(int(result['queries'].item()), queries)
def test_tensor_budget_cannot_wrap_depth(self):
a, b = torch.zeros(1, 6, dtype=torch.float64), torch.zeros(1, 6, dtype=torch.float64)
a[:, 0] = b[:, 0] = 0.25
for tensor in (torch.tensor(255, dtype=torch.uint8), torch.tensor(255)):
with self.subTest(dtype=tensor.dtype):
with self.assertRaisesRegex(ValueError, 'plain ints'):
self.ThinObstacle().certify_segments(a, b, max_depth=tensor)
def test_batch_matches_per_root_and_ignores_order(self):
kin = self.ThinObstacle()
a = torch.tensor([[0.0, 0, 0, 0, 0, 0], [0.25, 0, 0, 0, 0, 0], [-1.0, 0, 0, 0, 0, 0]], dtype=torch.float64)
b = torch.tensor([[1.0, 0, 0, 0, 0, 0], [0.25, 0, 0, 0, 0, 0], [0.0, 0, 0, 0, 0, 0]], dtype=torch.float64)
batch = kin.certify_segments(a, b, max_queries=16)
self.assertEqual(batch['collision'].tolist(), [True, True, False])
self.assertEqual(batch['safe'].tolist(), [False, False, True])
for perm in ([0, 1, 2], [2, 1, 0], [1, 0, 2]):
index = torch.tensor(perm)
moved = kin.certify_segments(a[index], b[index], max_queries=16)
for key in ('safe', 'collision', 'unresolved', 'queries'):
torch.testing.assert_close(moved[key], batch[key][index], rtol=0, atol=0)
def test_empty_batch_and_all_rejected_frontier(self):
kin = self.ThinObstacle()
empty = kin.certify_segments(torch.zeros(0, 6, dtype=torch.float64), torch.zeros(0, 6, dtype=torch.float64))
self.assertEqual(empty['safe'].shape, (0,))
a = torch.tensor([[0.25, 0, 0, 0, 0, 0], [6.0, 0, 0, 0, 0, 0]], dtype=torch.float64)
result = kin.certify_segments(a, a.clone(), max_queries=2)
self.assertEqual(result['safe'].tolist(), [False, False])
self.assertEqual(result['collision'].tolist(), [True, False])
self.assertEqual(result['outside_box'].tolist(), [False, True])
exp: 轮次32——均匀细分认证一次调用可证段内碰撞，界经720项检验未证伪，但P50≤20ms的档只能证出54~63%
2026-10-20 19:30:02 +08:00
def test_uniform_matches_bound_formula_and_needs_no_round_trips(self):
class ConstantGap(self.ThinObstacle):
def gaps(self, q, cap):
return {'gm': torch.full((q.shape[0], 1), 0.1, dtype=q.dtype, device=q.device)}
# 界 1 mm/度、间隙 0.1 mm：子段判据 0.2 > 1/n 需要 n ≥ 6，n=5 必须落 unproven。
kin = ConstantGap()
kin.cert_lever = torch.ones(1, 6, dtype=torch.float64)
a, b = torch.zeros(1, 6, dtype=torch.float64), torch.zeros(1, 6, dtype=torch.float64)
b[:, 0] = 1.0
for n, safe in ((5, False), (6, True)):
with self.subTest(n=n):
result = kin.certify_uniform(a, b, n=n)
self.assertEqual(bool(result['safe'].item()), safe)
self.assertEqual(result['points'], n + 1)
self.assertFalse(bool(result['collision'].item()))
self.assertEqual(bool(result['unproven'].item()), not safe)
self.assertAlmostEqual(result['cap_required_mm'], 1.0 / n + 2e-4, places=9)
def test_uniform_catches_hidden_collision_and_rejects_short_grids(self):
a = torch.tensor([[0.0, 0, 0, 0, 0, 0], [-1.0, 0, 0, 0, 0, 0], [6.0, 0, 0, 0, 0, 0]], dtype=torch.float64)
b = torch.tensor([[1.0, 0, 0, 0, 0, 0], [0.0, 0, 0, 0, 0, 0], [6.0, 0, 0, 0, 0, 0]], dtype=torch.float64)
kin = self.ThinObstacle()
short = kin.certify_uniform(a, b, n=2)
self.assertEqual(short['collision'].tolist(), [False, False, False])
self.assertEqual(short['unproven'].tolist(), [True, False, False])
self.assertEqual(short['safe'].tolist(), [False, True, False])
for n in (4, 8, 64):
with self.subTest(n=n):
result = kin.certify_uniform(a, b, n=n)
self.assertEqual(result['collision'].tolist(), [True, False, False])
self.assertEqual(result['safe'].tolist(), [False, True, False])
self.assertEqual(result['outside_box'].tolist(), [False, False, True])
def test_uniform_requires_plain_int_count(self):
a = torch.zeros(1, 6, dtype=torch.float64)
for tensor in (torch.tensor(8, dtype=torch.uint8), torch.tensor(8)):
with self.subTest(dtype=tensor.dtype):
with self.assertRaisesRegex(ValueError, 'subdivision count'):
self.ThinObstacle().certify_uniform(a, a.clone(), n=tensor)
exp+fix: 轮次31——确定界连续认证抓到段内碰撞，但单段P50 22.9ms进不了20ms；自适应cap收益经交错复测撤回
2026-10-20 19:13:40 +08:00
def test_nonfinite_and_outside_box_are_rejected(self):
a, b = torch.zeros(1, 6, dtype=torch.float64), torch.zeros(1, 6, dtype=torch.float64)
kin = self.ThinObstacle()
with patch.object(kin, 'gaps', side_effect=lambda q, cap: {'gm': torch.full((q.shape[0], 1), float('nan'))}):
result = kin.certify_segments(a, b)
self.assertFalse(result['safe'].item())
self.assertTrue(result['unresolved'].item())
for value in (float('inf'), -float('inf')):
with self.subTest(value=value):
with patch.object(kin, 'gaps',
side_effect=lambda q, cap, v=value: {'gm': torch.full((q.shape[0], 1), v)}):
result = kin.certify_segments(a, b)
self.assertFalse(result['safe'].item())
a[:, 0] = b[:, 0] = 6.0
result = kin.certify_segments(a, b)
self.assertFalse(result['safe'].item())
self.assertTrue(result['outside_box'].item())
def test_point_segment_distance_is_symmetric_and_differentiable(self):
p = torch.tensor([[0.0, 0, 0], [0, 0, 0], [0, 0, 0], [0, 0, 0]], dtype=torch.float64, requires_grad=True)
q = torch.tensor([[2.0, 0, 0]] * 4, dtype=torch.float64)
point = torch.tensor([[1.0, 0, 0], [1, 3, 0], [4, 0, 0], [-1, 0, 0]], dtype=torch.float64)
forward = seg_seg_dist(p, q, point, point)
reverse = seg_seg_dist(point, point, p, q)
torch.testing.assert_close(forward, torch.tensor([0.0, 3, 2, 1], dtype=torch.float64), atol=1e-8, rtol=0)
torch.testing.assert_close(forward, reverse, atol=1e-8, rtol=0)
forward.sum().backward()
self.assertTrue(torch.isfinite(p.grad).all())
torch.testing.assert_close(seg_seg_dist(p.detach(), p.detach(), point, point),
(p.detach() - point).norm(dim=-1), atol=1e-8, rtol=0)
exp: 轮次33(上)——实际执行速度做成一等指标并自证；默认路径逐位复现第30轮，256题全部超VMAX
2026-10-20 19:48:23 +08:00
class SpeedAndMarginTest(unittest.TestCase):
"""轮次33：反馈叠加后的每轴速度限幅，以及安全层的正余量门限（含 cap 截顶坑）。"""
STEP = [1.2, 1.2, 1.8, 1.8, 2.4, 3.6]
def rollout_one(self, teacher_step, gap_value, fb=None, **kwargs):
class FlatKin:
def clamp_box(self, q):
return q
def gaps(self, q, cap):
return {'gm': torch.full((q.shape[0], 49), float(gap_value))}
commit_n = 2
traj = torch.zeros(1, commit_n + 1, 6)
traj[0, 1:] = torch.tensor(teacher_step)[None].repeat(commit_n, 1).cumsum(0) # teacher_step = 每步各轴的度数增量
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])),
}
with patch.multiple('ml.sim.closed_loop', **patches):
with patch('ml.sim.closed_loop.dls_steps', return_value=torch.tensor([[[1.0, 0, 0, 0, 0, 0]]])):
_, per = rollout(None, FlatKin(), [{'q0': [0] * 6, 'T': [10, 0, 0]}], 'cpu', K=1, safety=True,
max_sec=0.02 * commit_n, return_traj=True, use_sub=False, select='first',
oracle=TeacherOracle(traj), replan='empty', commit_n=commit_n,
**kwargs)
return per
def test_feedback_over_speed_is_measured_not_assumed(self):
# 策略给满 STEP（J1 +1.2°）+ 2.2×1.0° 反馈 = 3.4° > VMAX·DT
per = self.rollout_one([1.2, 0, 0, 0, 0, 0], 5.0, fb_gain=2.2)
self.assertAlmostEqual(per['traj'][0, 1, 0].item(), 3.4, places=5)
self.assertAlmostEqual(per['speed_ratio'][0].item(), 3.4 / 1.2, places=5)
def test_inc_cap_limits_actual_step_to_vmax(self):
per = self.rollout_one([1.2, 0, 0, 0, 0, 0], 5.0, fb_gain=2.2, inc_cap=1)
self.assertAlmostEqual(per['traj'][0, 1, 0].item(), 1.2, places=5)
self.assertLessEqual(per['speed_ratio'][0].item(), 1.0 + 1e-6)
def test_margin_rejects_near_obstacle_step_that_passes_at_zero(self):
moved = self.rollout_one([0.5, 0, 0, 0, 0, 0], 0.3, safety_partial=1)
self.assertAlmostEqual(moved['traj'][0, -1, 0].item(), 1.0, places=5)
self.assertEqual(moved['brakes'].item(), 0.0)
braked = self.rollout_one([0.5, 0, 0, 0, 0, 0], 0.3, gap_margin=0.5, safety_partial=1)
self.assertEqual(braked['brakes'].item(), 1.0)
torch.testing.assert_close(braked['traj'], torch.zeros(1, 3, 6), rtol=0, atol=0)
def test_margin_below_cap_still_accepts_clear_step(self):
# 余量 0.5 必须仍能通过 5 mm 间隙：若宽相位 cap 没抬到 margin+0.5，截顶值 0.5 会让门限恒假
per = self.rollout_one([0.5, 0, 0, 0, 0, 0], 5.0, gap_margin=0.5, safety_partial=1)
self.assertEqual(per['brakes'].item(), 0.0)
self.assertAlmostEqual(per['traj'][0, -1, 0].item(), 1.0, places=5)
self.assertTrue((per['plans']['gaps'] > 0.5).all().item())
perf+exp: 轮次34——复用cum_xf进库(22项指标逐位不变)、批量认证压最坏尾部(80→37ms)、FP16/BF16实测无收益
2026-10-20 20:53:43 +08:00
class FeatureGeometryReuseTest(unittest.TestCase):
"""#16 第①条：gm_groups 复用调用者已算好的 cum_xf，省一次 FK，但结果必须与原口径逐位一致。"""
def test_gm_groups_with_supplied_xf_is_bitwise_identical(self):
from ml.common.features import features, GP, GAP_CAP
kin = Kin()
q = torch.tensor([[-129.9, 46.68, 43.49, -97.46, -85.1, -139.7],
[10.0, -20.0, 30.0, -40.0, 50.0, -60.0]])
xf = kin.cum_xf(q)
torch.testing.assert_close(kin.gm_groups(q, GP, GAP_CAP, xf=xf), kin.gm_groups(q, GP, GAP_CAP),
rtol=0, atol=0)
idx = [a * 7 + b for a, b in GP]
table = kin.gaps(q, GAP_CAP, xf)['gm'][:, idx]
torch.testing.assert_close(kin.gm_groups(q, GP, GAP_CAP), table, rtol=0, atol=0)
z = torch.zeros(len(q), 6)
f = features(kin, q, z, q[:, :3], torch.zeros(len(q), 9), z[:, None, :].expand(len(q), 3, 6),
torch.zeros(len(q)), z, torch.zeros(len(q)))
cols = 12 + 3 + 6 + 3 + 6 + 3 + 6 + 3 + 3 + 2 + 18 + 3 + 6 + 1 # 间隙块之前的特征数 = 75
self.assertEqual(cols + 8, 83)
torch.testing.assert_close(f[:, cols:cols + 8], table / GAP_CAP, rtol=0, atol=0)
class CertBatchEquivalenceTest(unittest.TestCase):
"""#16 第②条：把 K 条候选的认证并成一次 gaps 调用，选择结果必须与串行循环逐位一致。"""
class StubModel:
act_scale = 1.0
def __init__(self, act):
self.act = act
def sample_k(self, feat, K, z, steps=1):
B = feat.shape[0]
return (self.act.expand(B, K, 16, 6).clone(), torch.zeros(B, device=feat.device),
torch.zeros(B, 13, device=feat.device))
def run_both(self, acts, reject_first):
class StepKin:
def clamp_box(self, q):
return q
def gaps(self, q, cap):
bad = (q[:, 0] > 0.5) if reject_first else torch.zeros(len(q), dtype=torch.bool, device=q.device)
return {'gm': torch.where(bad[:, None], -torch.ones(len(q), 49, device=q.device),
torch.full((len(q), 49), 1.0, device=q.device))}
acts = torch.tensor(acts, dtype=torch.float32) # [K,16,6] 动作（±1 归一化）
model = self.StubModel(acts)
tasks = [{'q0': [0] * 6, 'T': [10, 0, 0]}] * 2 # B=2：批量索引形状必须在多行下也成立
out = []
for batched in (0, 1):
with patch('ml.sim.closed_loop.features', return_value=torch.zeros(len(tasks), 83)), \
patch('ml.sim.closed_loop.goal_err',
side_effect=lambda kin, q, T, R: (torch.full((q.shape[0],), 10.0), torch.zeros(q.shape[0]))):
_, per = rollout(model, StepKin(), tasks, 'cpu', K=acts.shape[0], safety=True, max_sec=0.04,
return_traj=True, use_sub=False, select='first', replan='empty', commit_n=2,
safety_partial=1, cert_batch=batched)
out.append(per)
return out
def test_batched_certification_selects_the_same_candidate(self):
# 候选 0 第 2 步撞上 q0>0.5 的障碍 ⇒ 只有前缀 1 步；候选 1 全程安全 ⇒ 应被选中
acts = [[[0.2, 0, 0, 0, 0, 0], [0.5, 0, 0, 0, 0, 0]] + [[0, 0, 0, 0, 0, 0]] * 14,
[[0.1, 0, 0, 0, 0, 0], [0.1, 0, 0, 0, 0, 0]] + [[0, 0, 0, 0, 0, 0]] * 14]
seq, batched = self.run_both(acts, reject_first=True)
torch.testing.assert_close(batched['traj'], seq['traj'], rtol=0, atol=0)
torch.testing.assert_close(batched['plans']['commit'], seq['plans']['commit'], rtol=0, atol=0)
torch.testing.assert_close(batched['plans']['certified'], seq['plans']['certified'], rtol=0, atol=0)
self.assertEqual(batched['brakes'].tolist(), seq['brakes'].tolist())
self.assertAlmostEqual(seq['traj'][0, -1, 0].item(), 0.24, places=5) # 候选 1 的两步 0.1×STEP=0.12°
def test_batched_certification_keeps_partial_prefix_choice(self):
# 所有候选都只走得了第 1 步 ⇒ 两条路径都取第 1 条候选、前缀长度 1
acts = [[[0.2, 0, 0, 0, 0, 0], [0.5, 0, 0, 0, 0, 0]] + [[0, 0, 0, 0, 0, 0]] * 14,
[[0.3, 0, 0, 0, 0, 0], [0.6, 0, 0, 0, 0, 0]] + [[0, 0, 0, 0, 0, 0]] * 14]
seq, batched = self.run_both(acts, reject_first=True)
torch.testing.assert_close(batched['traj'], seq['traj'], rtol=0, atol=0)
torch.testing.assert_close(batched['plans']['commit'], seq['plans']['commit'], rtol=0, atol=0)
self.assertAlmostEqual(seq['traj'][0, 1, 0].item(), 0.24, places=5)
self.assertAlmostEqual(seq['traj'][0, 2, 0].item(), 0.24, places=5) # 第 2 步被拒后停在原地
fix+exp: 轮次35(补)——gaps/gm_groups 拒绝行数不符的 xf；撤回两次不存在的"合并调用"收益
2026-10-20 21:06:35 +08:00
def test_mismatched_xf_row_count_is_rejected(self):
kin = Kin()
q1 = torch.zeros(1, 6); q5 = torch.zeros(5, 6)
xf5 = kin.cum_xf(q5)
for name, fn in (('gaps', lambda: kin.gaps(q1, 40.0, xf5)),
('gm_groups', lambda: kin.gm_groups(q1, [(0, 2)], 40.0, xf=xf5))):
with self.subTest(fn=name):
with self.assertRaisesRegex(ValueError, '行数'):
fn()
self.assertEqual(kin.gaps(q5, 40.0, xf5)['gm'].shape, (5, 49))
self.assertEqual(kin.gm_groups(q5, [(0, 2)], 40.0, xf=xf5).shape, (5, 1))
exp: 轮次29——全拒后减半反馈到达55.1%，补齐无碰撞指标与首撞回放
2026-10-20 17:06:53 +08:00
if __name__ == '__main__':
unittest.main()
Reference in New Issue Copy Permalink