Brain episodes on pymunk-reference, superdex-reference
Rendering matches the published pygame frame (white canvas, LightSlateGray T with its two-box seam, RoyalBlue pusher, LightGreen goal, red cross = the commanded target). 40 episodes from runs/episodes.sqlite; planner rule; executor at the measured human mouse-stream speeds; 200 steps at 10 Hz. Captions are the stored narratives (docs/episode_store.md).
The same brain on SuperDex contact
superdex-reference is the published Push-T task - same T, goal, seeded start, 100 Hz pixel PD law on a kinematic pusher - running on SuperDex physics: a real table with Coulomb friction (0.5) under the block, a penalty contact with friction (0.5) between pusher and block, and the block's geometric inertia, where the published environment has no friction anywhere and a velocity wipe for support. Same 10 seeds and the same rule planner as the pymunk-reference sections below; each episode is paired with the latest 200-step pymunk-reference episode on its seed.
Engine comparison. Same seeds, same planner, on pymunk-reference versus on superdex-reference.
seed
situation
bouts
peak px/s
angle error
position error
coverage
cov ≥ .95 at
SOLVED at
577
aligned, far
on pymunk-reference
3
114
3° → 1°
173 → 3 px
0.923
–
–
on superdex-reference
3
114
3° → 2°
173 → 40 px
0.377
–
–
428
aligned, moderate
on pymunk-reference
3
112
2° → 2°
59 → 1 px
0.954
66
70
on superdex-reference
4
114
2° → 1°
59 → 2 px
0.957
127
131
412
inverted, near
on pymunk-reference
5
114
167° → 75°
32 → 9 px
0.298
–
–
on superdex-reference
5
113
167° → 31°
32 → 15 px
0.469
–
–
422
mixed, far
on pymunk-reference
5
114
75° → 1°
103 → 2 px
0.961
192
195
on superdex-reference
6
114
75° → 4°
103 → 1 px
0.913
–
–
438
mixed, moderate
on pymunk-reference
6
114
41° → 4°
82 → 0 px
0.926
–
–
on superdex-reference
4
114
41° → 3°
82 → 30 px
0.338
–
–
402
flipped, far, near a wall
on pymunk-reference
2
114
138° → 130°
142 → 136 px
0.000
–
–
on superdex-reference
2
114
138° → 128°
142 → 142 px
0.000
–
–
430
flipped, moderate
on pymunk-reference
4
114
126° → 18°
97 → 18 px
0.511
–
–
on superdex-reference
4
114
126° → 13°
97 → 33 px
0.254
–
–
429
mixed, far
on pymunk-reference
3
114
24° → 2°
186 → 3 px
0.933
–
–
on superdex-reference
3
114
24° → 2°
186 → 9 px
0.812
–
–
425
inverted, far, near a wall
on pymunk-reference
11
111
170° → 150°
150 → 147 px
0.000
–
–
on superdex-reference
11
111
170° → 142°
150 → 153 px
0.000
–
–
787
already nearly solved
on pymunk-reference
4
114
19° → 5°
18 → 1 px
0.901
–
–
on superdex-reference
2
114
19° → 1°
18 → 1 px
0.977
71
75
2 of 10 solve outright (428, 787) on superdex-reference, against 2 of 10 on pymunk-reference. Final coverage is higher than on pymunk-reference on 3, unchanged on 2 (402, 425), lower on 5 (577, 422, 438, 430, 429).
577 · aligned, far. plans: translate → translate → translate → translate. 173 → 40 px, 3° → 2° in 3 bouts; coverage 0.377 when the horizon runs out at step 200. Plan 3 (translate) overshot: 13 → 40 px.On pymunk-reference (brain_pymunk_ref, 2026-09-25): 3 bouts, 3° → 1°, 173 → 3 px, coverage 0.923, not solved. On superdex-reference: 3 bouts, 3° → 2°, 173 → 40 px, coverage 0.377, not solved. Position error +37 px, angle +1°, coverage -0.546; still short of 0.95.cebc8fc04320428 · aligned, moderate. plans: translate → finish → finish. SOLVED at step 131: coverage crossed 0.95 at 127 and the pusher was 10 px clear 4 steps later; 59 → 2 px, 2° → 1°, coverage 0.957 in 4 bouts.On pymunk-reference (brain_pymunk_ref, 2026-09-25): 3 bouts, 2° → 2°, 59 → 1 px, coverage 0.954, SOLVED at step 70. On superdex-reference: 4 bouts, 2° → 1°, 59 → 2 px, coverage 0.957, SOLVED at step 131. Solved both times, at step 70 on pymunk-reference and 131 on superdex-reference.29664a029923412 · inverted, near. plans: rotate → rotate → rotate → rotate → rotate → rotate. 32 → 15 px, 167° → 31° in 5 bouts; coverage 0.469 when the horizon runs out at step 200. Plan 4 (rotate) overshot: 10 → 35 px.On pymunk-reference (brain_pymunk_ref, 2026-09-25): 5 bouts, 167° → 75°, 32 → 9 px, coverage 0.298, not solved. On superdex-reference: 5 bouts, 167° → 31°, 32 → 15 px, coverage 0.469, not solved. Position error +6 px, angle -44°, coverage +0.171; still short of 0.95.028374b78684422 · mixed, far. plans: translate → translate → translate → rotate → finish → finish. 2 finish plans in a row bring the position error to 1 px, but the angle sits at 4° and coverage stalls at 0.913: 0.95 needs ~2°.On pymunk-reference (brain_pymunk_ref, 2026-09-25): 5 bouts, 75° → 1°, 103 → 2 px, coverage 0.961, SOLVED at step 195. On superdex-reference: 6 bouts, 75° → 4°, 103 → 1 px, coverage 0.913, not solved. Position error -1 px, angle +4°, coverage -0.048; still short of 0.95.c2649d382026438 · mixed, moderate. plans: translate → rotate → translate → rotate → translate. 82 → 30 px, 41° → 3° in 4 bouts; coverage 0.338 when the horizon runs out at step 200. Plan 2 (rotate) overshot: 2 → 32 px.On pymunk-reference (brain_pymunk_ref, 2026-09-25): 6 bouts, 41° → 4°, 82 → 0 px, coverage 0.926, not solved. On superdex-reference: 4 bouts, 41° → 3°, 82 → 30 px, coverage 0.338, not solved. Position error +30 px, angle -1°, coverage -0.589; still short of 0.95.61303b3d0047402 · flipped, far, near a wall. plans: rotate → translate → translate. Little progress per bout: 2 bouts take 138° → 128° and 142 → 142 px (5.0° and 0.2 px per bout against the human's 5.5° and 20 px). The block starts 102 px from a wall (under 115): the likely cause is an edge that faces the wall.On pymunk-reference (brain_pymunk_ref, 2026-09-25): 2 bouts, 138° → 130°, 142 → 136 px, coverage 0.000, not solved. On superdex-reference: 2 bouts, 138° → 128°, 142 → 142 px, coverage 0.000, not solved. Position error +5 px, angle -2°, coverage 0.000; still short of 0.95.bbf5baa7391a430 · flipped, moderate. plans: rotate → rotate → rotate → translate. 97 → 33 px, 126° → 13° in 4 bouts; coverage 0.254 when the horizon runs out at step 200. Plan 3 (rotate) overshot: 10 → 36 px.On pymunk-reference (brain_pymunk_ref, 2026-09-25): 4 bouts, 126° → 18°, 97 → 18 px, coverage 0.511, not solved. On superdex-reference: 4 bouts, 126° → 13°, 97 → 33 px, coverage 0.254, not solved. Position error +16 px, angle -6°, coverage -0.257; still short of 0.95.23eb26bb7a5c429 · mixed, far. plans: translate → translate → translate → finish. 186 → 9 px, 24° → 2° in 3 bouts; coverage 0.812 when the horizon runs out at step 200.On pymunk-reference (brain_pymunk_ref, 2026-09-25): 3 bouts, 24° → 2°, 186 → 3 px, coverage 0.933, not solved. On superdex-reference: 3 bouts, 24° → 2°, 186 → 9 px, coverage 0.812, not solved. Position error +6 px, angle +1°, coverage -0.121; still short of 0.95.3df7d0dcc197425 · inverted, far, near a wall. plans: rotate → rotate → rotate → rotate → rotate → rotate → rotate → rotate → rotate → rotate → translate → translate. Little progress per bout: 11 bouts take 170° → 142° and 150 → 153 px (2.5° and -0.3 px per bout against the human's 5.5° and 20 px). The block starts 99 px from a wall (under 115): the likely cause is an edge that faces the wall.On pymunk-reference (brain_pymunk_ref, 2026-09-25): 11 bouts, 170° → 150°, 150 → 147 px, coverage 0.000, not solved. On superdex-reference: 11 bouts, 170° → 142°, 150 → 153 px, coverage 0.000, not solved. Position error +6 px, angle -7°, coverage 0.000; still short of 0.95.3133015a89ee787 · already nearly solved. plans: rotate → rotate. SOLVED at step 75: coverage crossed 0.95 at 71 and the pusher was 10 px clear 4 steps later; 18 → 1 px, 19° → 1°, coverage 0.977 in 2 bouts. The last plan was a rotate, not a finish: it solved on precision alone.On pymunk-reference (brain_pymunk_ref, 2026-09-25): 4 bouts, 19° → 5°, 18 → 1 px, coverage 0.901, not solved. On superdex-reference: 2 bouts, 19° → 1°, 18 → 1 px, coverage 0.977, SOLVED at step 75. Now SOLVED at step 75; on pymunk-reference it was not.498400a72fc1
The pymunk-reference baseline for the engine comparison
The same 10 seeds re-run on pymunk-reference with the code that produced the brain_superdex_ref episodes, so the engine comparison pairs like with like; the before/after rows are against each seed's earlier pymunk-reference episode.
seed
situation
bouts
peak px/s
angle error
position error
coverage
cov ≥ .95 at
SOLVED at
577
aligned, far
before
3
114
3° → 1°
173 → 38 px
0.360
–
–
after
3
114
3° → 1°
173 → 3 px
0.923
–
–
428
aligned, moderate
before
3
112
2° → 2°
59 → 1 px
0.954
66
70
after
3
112
2° → 2°
59 → 1 px
0.954
66
70
412
inverted, near
before
5
114
167° → 75°
32 → 9 px
0.298
–
–
after
5
114
167° → 75°
32 → 9 px
0.298
–
–
422
mixed, far
before
5
114
75° → 1°
103 → 2 px
0.961
192
195
after
5
114
75° → 1°
103 → 2 px
0.961
192
195
438
mixed, moderate
before
6
114
41° → 4°
82 → 0 px
0.926
–
–
after
6
114
41° → 4°
82 → 0 px
0.926
–
–
402
flipped, far, near a wall
before
2
114
138° → 130°
142 → 136 px
0.000
–
–
after
2
114
138° → 130°
142 → 136 px
0.000
–
–
430
flipped, moderate
before
4
114
126° → 18°
97 → 18 px
0.511
–
–
after
4
114
126° → 18°
97 → 18 px
0.511
–
–
429
mixed, far
before
3
114
24° → 0°
186 → 35 px
0.391
–
–
after
3
114
24° → 2°
186 → 3 px
0.933
–
–
425
inverted, far, near a wall
before
11
111
170° → 150°
150 → 147 px
0.000
–
–
after
11
111
170° → 150°
150 → 147 px
0.000
–
–
787
already nearly solved
before
4
114
19° → 5°
18 → 1 px
0.901
–
–
after
4
114
19° → 5°
18 → 1 px
0.901
–
–
2 of 10 solve outright (428, 422). Final coverage is higher than before on 2, unchanged on 8 (428, 412, 422, 438, 402, 430, 425, 787), lower on 0.
577 · aligned, far. plans: translate → translate → finish → finish. 173 → 3 px, 3° → 1° in 3 bouts; coverage 0.923 when the horizon runs out at step 200.Before (brain_rule, 2026-09-25): 3 bouts, 3° → 1°, 173 → 38 px, coverage 0.360, not solved. After: 3 bouts, 3° → 1°, 173 → 3 px, coverage 0.923, not solved. Position error -36 px, angle 0°, coverage +0.563; still short of 0.95.51443cb0cae8428 · aligned, moderate. plans: translate → finish. SOLVED at step 70: coverage crossed 0.95 at 66 and the pusher was 10 px clear 4 steps later; 59 → 1 px, 2° → 2°, coverage 0.954 in 3 bouts.Before (brain_finish, 2026-09-25): 3 bouts, 2° → 2°, 59 → 1 px, coverage 0.954, SOLVED at step 70. After: 3 bouts, 2° → 2°, 59 → 1 px, coverage 0.954, SOLVED at step 70. Identical to before, to the pixel.560477a27c1d412 · inverted, near. plans: rotate → rotate → rotate → rotate → rotate → rotate. 32 → 9 px, 167° → 75° in 5 bouts; coverage 0.298 when the horizon runs out at step 200. Plan 3 (rotate) overshot: 11 → 38 px.Before (brain_rule, 2026-09-25): 5 bouts, 167° → 75°, 32 → 9 px, coverage 0.298, not solved. After: 5 bouts, 167° → 75°, 32 → 9 px, coverage 0.298, not solved. Identical to before, to the pixel.80212882aa18422 · mixed, far. plans: translate → translate → rotate → translate. SOLVED at step 195: coverage crossed 0.95 at 192 and the pusher was 10 px clear 3 steps later; 103 → 2 px, 75° → 1°, coverage 0.961 in 5 bouts. The last plan was a translate, not a finish: it solved on precision alone.Before (brain_finish, 2026-09-25): 5 bouts, 75° → 1°, 103 → 2 px, coverage 0.961, SOLVED at step 195. After: 5 bouts, 75° → 1°, 103 → 2 px, coverage 0.961, SOLVED at step 195. Identical to before, to the pixel.7dfd49ee5908438 · mixed, moderate. plans: translate → translate → finish → finish → finish → finish → finish. 5 finish plans in a row bring the position error to 0 px, but the angle sits at 4° and coverage stalls at 0.926: 0.95 needs ~2°.Before (brain_finish, 2026-09-25): 6 bouts, 41° → 4°, 82 → 0 px, coverage 0.926, not solved. After: 6 bouts, 41° → 4°, 82 → 0 px, coverage 0.926, not solved. Identical to before, to the pixel.58e4492de34f402 · flipped, far, near a wall. plans: rotate → rotate → translate. Little progress per bout: 2 bouts take 138° → 130° and 142 → 136 px (4.2° and 2.8 px per bout against the human's 5.5° and 20 px). The block starts 102 px from a wall (under 115): the likely cause is an edge that faces the wall.Before (brain_rule, 2026-09-25): 2 bouts, 138° → 130°, 142 → 136 px, coverage 0.000, not solved. After: 2 bouts, 138° → 130°, 142 → 136 px, coverage 0.000, not solved. Identical to before, to the pixel.680c2112f194430 · flipped, moderate. plans: rotate → rotate → rotate → rotate → rotate. 97 → 18 px, 126° → 18° in 4 bouts; coverage 0.511 when the horizon runs out at step 200. Plan 3 (rotate) overshot: 11 → 37 px.Before (brain_finish, 2026-09-25): 4 bouts, 126° → 18°, 97 → 18 px, coverage 0.511, not solved. After: 4 bouts, 126° → 18°, 97 → 18 px, coverage 0.511, not solved. Identical to before, to the pixel.12c27103b7b1429 · mixed, far. plans: translate → translate → finish → finish. 186 → 3 px, 24° → 2° in 3 bouts; coverage 0.933 when the horizon runs out at step 200.Before (brain_rule, 2026-09-25): 3 bouts, 24° → 0°, 186 → 35 px, coverage 0.391, not solved. After: 3 bouts, 24° → 2°, 186 → 3 px, coverage 0.933, not solved. Position error -32 px, angle +2°, coverage +0.542; still short of 0.95.4fb86590d35b425 · inverted, far, near a wall. plans: rotate → rotate → rotate → rotate → rotate → rotate → rotate → rotate → rotate → rotate → rotate → rotate. Little progress per bout: 11 bouts take 170° → 150° and 150 → 147 px (1.9° and 0.3 px per bout against the human's 5.5° and 20 px). The block starts 99 px from a wall (under 115): the likely cause is an edge that faces the wall.Before (brain_rule, 2026-09-25): 11 bouts, 170° → 150°, 150 → 147 px, coverage 0.000, not solved. After: 11 bouts, 170° → 150°, 150 → 147 px, coverage 0.000, not solved. Identical to before, to the pixel.34c8ff2f8a26787 · already nearly solved. plans: rotate → rotate → translate → finish → finish. 2 finish plans in a row bring the position error to 1 px, but the angle sits at 5° and coverage stalls at 0.901: 0.95 needs ~2°. Plan 2 (rotate) overshot: 2 → 31 px.Before (brain_finish, 2026-09-25): 4 bouts, 19° → 5°, 18 → 1 px, coverage 0.901, not solved. After: 4 bouts, 19° → 5°, 18 → 1 px, coverage 0.901, not solved. Identical to before, to the pixel.86d061bfe80f
With a success criterion and a finishing behaviour
Solved = block coverage ≥ 0.95 and the pusher backed off ≥ 10 px from the T. The brain is told that finish line and the precision it needs (~6 px, ~2°, from the human final bouts); plans carry an intent (rotate / translate / finish); finish pushes through the CoG with a step scaled to the remaining error; and once coverage crosses 0.95 the pusher backs straight out along the outward normal to the 37 px standoff instead of idling in contact. Same 5 seeds, before and after.
seed
situation
bouts
peak px/s
angle error
position error
coverage
cov ≥ .95 at
SOLVED at
422
mixed, far
before
5
114
75° → 1°
103 → 3 px
0.944
192
–
after
5
114
75° → 1°
103 → 2 px
0.961
192
195
438
mixed, moderate
before
4
113
41° → 4°
82 → 2 px
0.909
–
–
after
6
114
41° → 4°
82 → 0 px
0.926
–
–
787
already nearly solved
before
4
114
19° → 3°
18 → 31 px
0.411
–
–
after
4
114
19° → 5°
18 → 1 px
0.901
–
–
430
flipped, moderate
before
4
114
126° → 18°
97 → 18 px
0.511
–
–
after
4
114
126° → 18°
97 → 18 px
0.511
–
–
428
aligned, moderate
before
5
112
2° → 4°
59 → 28 px
0.447
66
–
after
3
112
2° → 2°
59 → 1 px
0.954
66
70
2 of 5 solve outright (422, 428), none of which did before. Final coverage is higher than before on 4, unchanged on 1 (430), lower on 0.
422 · mixed, far. plans: translate → translate → rotate → translate. SOLVED at step 195: coverage crossed 0.95 at 192 and the pusher was 10 px clear 3 steps later; 103 → 2 px, 75° → 1°, coverage 0.961 in 5 bouts. The last plan was a translate, not a finish: it solved on precision alone.Before (brain_rule, 2026-09-25): 5 bouts, 75° → 1°, 103 → 3 px, coverage 0.944, coverage crossed 0.95 at step 192 but not solved. After: 5 bouts, 75° → 1°, 103 → 2 px, coverage 0.961, SOLVED at step 195. Now SOLVED at step 195; before it was not.a8d6e4ad885c438 · mixed, moderate. plans: translate → translate → finish → finish → finish → finish → finish. 5 finish plans in a row bring the position error to 0 px, but the angle sits at 4° and coverage stalls at 0.926: 0.95 needs ~2°.Before (brain_rule, 2026-09-25): 4 bouts, 41° → 4°, 82 → 2 px, coverage 0.909, not solved. After: 6 bouts, 41° → 4°, 82 → 0 px, coverage 0.926, not solved. Position error -1 px, angle 0°, coverage +0.018; still short of 0.95.9a5daa8adf08787 · already nearly solved. plans: rotate → rotate → translate → finish → finish. 2 finish plans in a row bring the position error to 1 px, but the angle sits at 5° and coverage stalls at 0.901: 0.95 needs ~2°. Plan 2 (rotate) overshot: 2 → 31 px.Before (brain_rule, 2026-09-25): 4 bouts, 19° → 3°, 18 → 31 px, coverage 0.411, not solved. After: 4 bouts, 19° → 5°, 18 → 1 px, coverage 0.901, not solved. The overshoot is gone: 18 → 31 px before, 18 → 1 px after.c4413d2d92f7430 · flipped, moderate. plans: rotate → rotate → rotate → rotate → rotate. 97 → 18 px, 126° → 18° in 4 bouts; coverage 0.511 when the horizon runs out at step 200. Plan 3 (rotate) overshot: 11 → 37 px.Before (brain_rule, 2026-09-25): 4 bouts, 126° → 18°, 97 → 18 px, coverage 0.511, not solved. After: 4 bouts, 126° → 18°, 97 → 18 px, coverage 0.511, not solved. Identical to before, to the pixel.43db96745302428 · aligned, moderate. plans: translate → finish. SOLVED at step 70: coverage crossed 0.95 at 66 and the pusher was 10 px clear 4 steps later; 59 → 1 px, 2° → 2°, coverage 0.954 in 3 bouts.Before (brain_rule, 2026-09-25): 5 bouts, 2° → 4°, 59 → 28 px, coverage 0.447, coverage crossed 0.95 at step 66 but not solved. After: 3 bouts, 2° → 2°, 59 → 1 px, coverage 0.954, SOLVED at step 70. Now SOLVED at step 70; before it was not.69a6d69527c3
15 episodes across the situation space
seed
situation
bouts
peak px/s
angle error
position error
coverage
577
aligned, far
3
114
3° → 1°
173 → 38 px
0.360
428
aligned, moderate
5
112
2° → 4°
59 → 28 px
0.447
412
inverted, near
5
114
167° → 75°
32 → 9 px
0.298
422
mixed, far
5
114
75° → 1°
103 → 3 px
0.944
438
mixed, moderate
4
113
41° → 4°
82 → 2 px
0.909
402
flipped, far, near a wall
2
114
138° → 130°
142 → 136 px
0.000
430
flipped, moderate
4
114
126° → 18°
97 → 18 px
0.511
429
mixed, far
3
114
24° → 0°
186 → 35 px
0.391
425
inverted, far, near a wall
11
111
170° → 150°
150 → 147 px
0.000
787
already nearly solved
4
114
19° → 3°
18 → 31 px
0.411
783958
inverted, moderate
5
113
163° → 47°
53 → 31 px
0.372
204
aligned, far
4
114
2° → 4°
175 → 9 px
0.778
234
flipped, near
4
114
124° → 27°
33 → 38 px
0.279
205
mixed, moderate
4
114
71° → 2°
83 → 10 px
0.792
194
flipped, far, near a wall
2
114
130° → 127°
129 → 127 px
0.000
Mean coverage 0.433; 5 of 15 above 0.5; 2 of 15 reach the 0.95 coverage line and 0 of 15 are solved in 200 steps. Peak pusher speed 111–114 px/s.
577 · aligned, far. plans: bar_top → stem_L → bar_top → stem_tip. 173 → 38 px, 3° → 1° in 3 bouts; coverage 0.360 when the horizon runs out at step 200. Plan 3 (bar_top) overshot: 11 → 38 px. 814a2c60c749428 · aligned, moderate. plans: stem_tip → stem_L → stem_R → stem_L → stem_R. Coverage crossed 0.95 at step 66 but no step had the pusher 10 px clear while it held, so it never counts as solved; final coverage 0.447 (max 0.955). Plan 2 (stem_L) overshot: 7 → 32 px. 68b3131bba3f412 · inverted, near. plans: bar_bottom_L → bar_end_L → bar_bottom_L → bar_top → bar_end_L → bar_top. 32 → 9 px, 167° → 75° in 5 bouts; coverage 0.298 when the horizon runs out at step 200. Plan 3 (bar_bottom_L) overshot: 11 → 38 px. 6239891f7e28422 · mixed, far. plans: bar_end_L → bar_bottom_L → bar_top → stem_tip → bar_top. Coverage crossed 0.95 at step 192 but no step had the pusher 10 px clear while it held, so it never counts as solved; final coverage 0.944 (max 0.969). Plan 3 (bar_top) overshot: 3 → 28 px. 4961a8065d99438 · mixed, moderate. plans: bar_end_R → bar_top → stem_R → stem_L → stem_R. The plans bring the position error to 2 px, but the angle sits at 4° and coverage stalls at 0.909: 0.95 needs ~2°. Plan 3 (stem_R) overshot: 6 → 34 px. 93eadb4723e4402 · flipped, far, near a wall. plans: bar_top → bar_top → bar_top. Little progress per bout: 2 bouts take 138° → 130° and 142 → 136 px (4.2° and 2.8 px per bout against the human's 5.5° and 20 px). The block starts 102 px from a wall (under 115): the likely cause is an edge that faces the wall. 2bb043199659430 · flipped, moderate. plans: bar_end_R → bar_top → bar_end_R → stem_L → bar_top. 97 → 18 px, 126° → 18° in 4 bouts; coverage 0.511 when the horizon runs out at step 200. Plan 3 (bar_end_R) overshot: 11 → 37 px. ad2e48185dea429 · mixed, far. plans: bar_top → bar_top → stem_L → stem_R. 186 → 35 px, 24° → 0° in 3 bouts; coverage 0.391 when the horizon runs out at step 200. Plan 3 (stem_L) overshot: 10 → 35 px. 47eeac3076c8425 · inverted, far, near a wall. plans: bar_bottom_R → bar_bottom_R → bar_bottom_R → bar_bottom_R → bar_bottom_R → bar_bottom_R → bar_bottom_R → bar_bottom_R → bar_bottom_R → bar_bottom_R → bar_bottom_R → bar_bottom_R. Little progress per bout: 11 bouts take 170° → 150° and 150 → 147 px (1.9° and 0.3 px per bout against the human's 5.5° and 20 px). The block starts 99 px from a wall (under 115): the likely cause is an edge that faces the wall. 0ce5e4d27b0b787 · already nearly solved. plans: bar_end_L → stem_R → stem_L → stem_R. Overshoot: plan 2 (stem_R) took the position error from 2 to 31 px and the episode ends farther from the goal than it started, 18 → 31 px (19° → 3°). Coverage peaked at 0.905 before ending at 0.411. 9a0d643d4b92
Earlier episodes, rendered with the previous colour scheme
783958 · inverted, moderate. plans: bar_end_L → bar_bottom_L → bar_end_L → stem_R → bar_top. 53 → 31 px, 163° → 47° in 5 bouts; coverage 0.372 when the horizon runs out at step 200. Plan 3 (bar_end_L) overshot: 5 → 32 px. 7ed33cd31d33204 · aligned, far. plans: stem_R → bar_top → stem_R → bar_top. 175 → 9 px, 2° → 4° in 4 bouts; coverage 0.778 when the horizon runs out at step 200. Coverage peaked at 0.893 before ending at 0.778. 3aaa1b8bbe92234 · flipped, near. plans: bar_top → bar_end_L → stem_R → bar_top → bar_bottom_L. Overshoot: plan 2 (bar_end_L) took the position error from 6 to 31 px and the episode ends farther from the goal than it started, 33 → 38 px (124° → 27°). Coverage peaked at 0.463 before ending at 0.279. abc17ed663cb205 · mixed, moderate. plans: stem_L → bar_top → stem_L → stem_R. 83 → 10 px, 71° → 2° in 4 bouts; coverage 0.792 when the horizon runs out at step 200. Plan 3 (stem_L) overshot: 8 → 34 px. 98165505e18e194 · flipped, far, near a wall. plans: bar_top → bar_top → bar_top. No progress: 129 → 127 px, 130° → 127°, coverage 0.000 after 2 bouts. The block starts 103 px from a wall (under 115): the likely cause is an edge that faces the wall. d79c980d82e4