EST1 / EST1b results (CRITERIA_EST1.json sha 64a04444…, CRITERIA_EST1b.json sha bc72beb9…; seeds 1344-1347; 228 + 152 runs)
Boundary force (zero-step holds cross 50 %), capture-prediction release only, final controller information paths; entry = first control tick in the hold mode after push onset.
| direction |
arm |
boundary |
ratio to oracle |
entry median (IQR) |
predicted exp(-omega dL) |
verdict |
| fore-aft |
oracle |
54.2 N |
1.000 |
0 ms (0-0) |
1.000 |
- |
| fore-aft |
ours (IMU-resolved) |
51.7 N |
0.954 |
15 ms (15-15) |
0.945 |
PASS |
| fore-aft |
tuned MOB |
46.2 N |
0.854 |
40 ms (35-45) |
0.859 |
PASS |
| fore-aft |
enhanced MOB (Sun 2026) |
45.8 N |
0.846 |
10 ms (10-40) |
0.963 |
FAIL |
| fore-aft |
centroidal + IMU (Fedsi 2026) |
51.2 N |
0.946 |
15 ms (15-15) |
0.945 |
PASS |
| lateral |
oracle |
74.2 N |
1.000 |
0 ms (0-0) |
1.000 |
- |
| lateral |
ours (IMU-resolved) |
71.7 N |
0.966 |
10 ms (10-10) |
0.963 |
PASS |
| lateral |
tuned MOB |
65.0 N |
0.876 |
30 ms (30-35) |
0.893 |
PASS |
| lateral |
enhanced MOB (Sun 2026) |
62.5 N |
0.843 |
30 ms (20-30) |
0.893 |
PASS |
| lateral |
centroidal + IMU (Fedsi 2026) |
71.2 N |
0.961 |
10 ms (10-10) |
0.963 |
PASS |
- EST1 E1 (latency predicts the boundary, +-0.08): PASS 4/4, largest difference 0.017.
- EST1 E2 (ours > tuned MOB): PASS both directions (51.7 vs 46.3 N; 71.7 vs 65.0 N).
- EST1 E3 (reported): entry ours 15/10 ms, MOB 40/30 ms.
- EST1b E1b: FAIL (3 of 4): centroidal + IMU passes both directions (0.946 vs 0.945; 0.961 vs 0.963); enhanced MOB passes lateral (0.843 vs 0.893) and fails fore-aft (0.846 vs 0.963).
- Post-hoc diagnosis (not confirmatory): the enhanced MOB’s lead term produces a wrong-sign transient of -21 to -32 N about 10 ms after every fore-aft push onset. Below 47.5 N it stays under the 25 N entry threshold and the hold is entered at 35-50 ms (all held up to 42.5 N); from 47.5 N it crosses the threshold, the hold is entered at 10 ms in the wrong direction and the robot steps or falls. The registered entry latency (first hold tick) therefore counts a wrong-direction entry; with the time of the correct-direction entry (35-50 ms) the prediction would be about 0.86. This is an instance of push-band error (NZ1), not of latency.