File size: 12,466 Bytes
0c3dd1d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 | # Case 2 held-out evaluation on goal set B (Sim2Sim)
Generated by `report_confirm.py` from `results/confirm_analysis.json` (output of `analyze_confirm.py`).
Goal set B: 75 ID (10-20 m) + 75 OOD (20-40 m) goals, never used before the TR design was fixed.
Primary pool: 12 held-out PPO policies (seeds 2001-2004). Secondary pool: the 12 + released checkpoints 1-4.
Condition 'paper_tr' is the noise-only TR (action and powertrain noise).
Bootstrap: 5000 goal resamples, stratified by range, paired across methods. Ranges are 95 % percentile ranges.
**Primary test: delta tau = +0.515 (95 % range +0.33 to +0.73); met = True.**
## Agreement with Chrono, primary pool (12 policies)
| Condition | tau, all goals (95 % range) | Delta tau vs nominal (95 % range) | tau ID | tau OOD | Spearman | AUC for Chrono success < 0.5 (95 % range) | Resolved pairs | tau with primary TR order |
|---|---|---|---:|---:|---:|---|---:|---:|
| nominal | 0.21 (0.03 to 0.36) | +0.00 (+0.00 to +0.00) | 0.11 | 0.15 | 0.29 | nan (n/a) | 54/66 | 0.12 |
| tr_rate (primary) | 0.73 (0.61 to 0.82) | +0.52 (+0.33 to +0.73) | 0.24 | 0.64 | 0.83 | nan (n/a) | 60/66 | 1.00 |
| paper_tr | 0.15 (0.00 to 0.27) | -0.06 (-0.18 to +0.03) | 0.27 | 0.09 | 0.27 | nan (n/a) | 60/66 | 0.06 |
| tr_rate_half | 0.67 (0.58 to 0.79) | +0.45 (+0.27 to +0.70) | 0.21 | 0.73 | 0.78 | nan (n/a) | 59/66 | 0.88 |
| tr_rate_double | 0.76 (0.61 to 0.85) | +0.55 (+0.33 to +0.73) | 0.24 | 0.48 | 0.89 | nan (n/a) | 59/66 | 0.91 |
| tr_rate_uniform | 0.76 (0.61 to 0.85) | +0.55 (+0.36 to +0.73) | 0.27 | 0.58 | 0.89 | nan (n/a) | 62/66 | 0.91 |
| tr_delay | 0.70 (0.58 to 0.79) | +0.48 (+0.30 to +0.70) | 0.24 | 0.82 | 0.83 | nan (n/a) | 63/66 | 0.91 |
| det_rate_0.1 | 0.67 (0.55 to 0.79) | +0.45 (+0.24 to +0.70) | 0.15 | 0.76 | 0.78 | nan (n/a) | 55/66 | 0.88 |
| det_calibrated | 0.82 (0.64 to 0.91) | +0.61 (+0.42 to +0.70) | 0.58 | 0.55 | 0.92 | nan (n/a) | 56/66 | 0.67 |
Chrono resolved pairs: 48/66. Chrono success < 0.5: none.
Pairs ordered as Chrono orders them: primary TR 57/66, nominal ROM 40/66. On the 29 pairs where the two disagree, Chrono supports the TR on 23 and nominal ROM on 6.
### Per policy, primary pool (mean rank: 1 = best of 12)
| Policy | Training config | Seed | Mean step change of steering (nominal ROM) | Chrono success ID | Chrono success OOD | Chrono mean rank | Nominal ROM success ID / OOD | Nominal ROM mean rank | TR success ID / OOD | TR mean rank |
|---|---|---:|---:|---:|---:|---:|---|---:|---|---:|
| robust_s2003 | robust | 2003 | 0.026 | 0.64 | 0.97 | 4.27 | 1.00 / 1.00 | 7.78 | 0.98 / 1.00 | 4.27 |
| robust_s2004 | robust | 2004 | 0.031 | 0.63 | 0.96 | 5.22 | 1.00 / 1.00 | 5.94 | 1.00 / 1.00 | 3.34 |
| release_s2001 | release | 2001 | 0.250 | 0.65 | 0.96 | 5.82 | 1.00 / 1.00 | 5.72 | 0.87 / 0.85 | 5.92 |
| nonoise_s2002 | nonoise | 2002 | 0.458 | 0.64 | 0.95 | 6.24 | 1.00 / 1.00 | 5.97 | 0.83 / 0.80 | 6.72 |
| release_s2003 | release | 2003 | 0.067 | 0.63 | 0.91 | 6.34 | 1.00 / 1.00 | 5.18 | 0.80 / 0.80 | 6.22 |
| release_s2002 | release | 2002 | 0.446 | 0.63 | 0.87 | 6.58 | 1.00 / 1.00 | 5.85 | 0.80 / 0.77 | 7.09 |
| nonoise_s2001 | nonoise | 2001 | 0.285 | 0.63 | 0.83 | 6.64 | 1.00 / 1.00 | 5.55 | 0.79 / 0.77 | 7.05 |
| robust_s2001 | robust | 2001 | 0.307 | 0.64 | 0.80 | 6.87 | 1.00 / 1.00 | 5.64 | 0.80 / 0.78 | 7.16 |
| nonoise_s2003 | nonoise | 2003 | 0.040 | 0.60 | 0.93 | 6.95 | 1.00 / 1.00 | 7.32 | 0.95 / 0.97 | 5.21 |
| nonoise_s2004 | nonoise | 2004 | 0.273 | 0.65 | 0.85 | 7.01 | 1.00 / 1.00 | 6.79 | 0.80 / 0.80 | 7.22 |
| robust_s2002 | robust | 2002 | 0.881 | 0.64 | 0.80 | 7.72 | 1.00 / 1.00 | 6.85 | 0.79 / 0.74 | 8.39 |
| release_s2004 | release | 2004 | 0.925 | 0.60 | 0.67 | 8.33 | 1.00 / 1.00 | 9.40 | 0.78 / 0.73 | 9.42 |
Order by mean rank:
- Chrono: robust_s2003 > robust_s2004 > release_s2001 > nonoise_s2002 > release_s2003 > release_s2002 > nonoise_s2001 > robust_s2001 > nonoise_s2003 > nonoise_s2004 > robust_s2002 > release_s2004
- Nominal ROM: release_s2003 > nonoise_s2001 > robust_s2001 > release_s2001 > release_s2002 > robust_s2004 > nonoise_s2002 > nonoise_s2004 > robust_s2002 > nonoise_s2003 > robust_s2003 > release_s2004
- Primary TR: robust_s2004 > robust_s2003 > nonoise_s2003 > release_s2001 > release_s2003 > nonoise_s2002 > nonoise_s2001 > release_s2002 > robust_s2001 > nonoise_s2004 > robust_s2002 > release_s2004
Gap magnitude: Chrono time / nominal ROM time on goals both reach = 1.29 to 1.54 (median 1.41) across policies.
## Agreement with Chrono, secondary pool (16 policies)
| Condition | tau, all goals (95 % range) | Delta tau vs nominal (95 % range) | tau ID | tau OOD | Spearman | AUC for Chrono success < 0.5 (95 % range) | Resolved pairs | tau with primary TR order |
|---|---|---|---:|---:|---:|---|---:|---:|
| nominal | 0.55 (0.45 to 0.63) | +0.00 (+0.00 to +0.00) | 0.49 | 0.53 | 0.68 | nan (n/a) | 108/120 | 0.35 |
| tr_rate (primary) | 0.70 (0.52 to 0.75) | +0.15 (-0.03 to +0.25) | 0.47 | 0.42 | 0.88 | nan (n/a) | 101/120 | 1.00 |
| paper_tr | 0.50 (0.42 to 0.57) | -0.05 (-0.10 to +0.02) | 0.58 | 0.48 | 0.66 | nan (n/a) | 114/120 | 0.30 |
| tr_rate_half | 0.13 (0.10 to 0.23) | -0.42 (-0.50 to -0.25) | -0.05 | 0.10 | 0.14 | nan (n/a) | 109/120 | 0.43 |
| tr_rate_double | 0.87 (0.78 to 0.92) | +0.32 (+0.22 to +0.42) | 0.58 | 0.67 | 0.96 | nan (n/a) | 113/120 | 0.80 |
| tr_rate_uniform | 0.82 (0.75 to 0.88) | +0.27 (+0.17 to +0.38) | 0.58 | 0.65 | 0.93 | nan (n/a) | 115/120 | 0.85 |
| tr_delay | 0.52 (0.40 to 0.65) | -0.03 (-0.18 to +0.15) | 0.57 | 0.22 | 0.69 | nan (n/a) | 97/120 | 0.75 |
| det_rate_0.1 | 0.15 (0.08 to 0.22) | -0.40 (-0.52 to -0.25) | -0.08 | 0.10 | 0.14 | nan (n/a) | 103/120 | 0.42 |
| det_calibrated | 0.87 (0.77 to 0.92) | +0.32 (+0.23 to +0.38) | 0.78 | 0.70 | 0.95 | nan (n/a) | 110/120 | 0.63 |
Chrono resolved pairs: 102/120. Chrono success < 0.5: none.
Pairs ordered as Chrono orders them: primary TR 102/120, nominal ROM 93/120. On the 39 pairs where the two disagree, Chrono supports the TR on 24 and nominal ROM on 15.
### Per policy, secondary pool (mean rank: 1 = best of 16)
| Policy | Training config | Seed | Mean step change of steering (nominal ROM) | Chrono success ID | Chrono success OOD | Chrono mean rank | Nominal ROM success ID / OOD | Nominal ROM mean rank | TR success ID / OOD | TR mean rank |
|---|---|---:|---:|---:|---:|---:|---|---:|---|---:|
| robust_s2003 | robust | 2003 | 0.026 | 0.64 | 0.97 | 4.98 | 1.00 / 1.00 | 7.78 | 0.98 / 1.00 | 5.11 |
| robust_s2004 | robust | 2004 | 0.031 | 0.63 | 0.96 | 5.99 | 1.00 / 1.00 | 5.94 | 1.00 / 1.00 | 4.00 |
| release_s2001 | release | 2001 | 0.250 | 0.65 | 0.96 | 6.56 | 1.00 / 1.00 | 5.75 | 0.87 / 0.85 | 7.55 |
| nonoise_s2002 | nonoise | 2002 | 0.458 | 0.64 | 0.95 | 7.05 | 1.00 / 1.00 | 5.97 | 0.83 / 0.80 | 8.52 |
| release_s2003 | release | 2003 | 0.067 | 0.63 | 0.91 | 7.28 | 1.00 / 1.00 | 5.18 | 0.80 / 0.80 | 7.85 |
| release_s2002 | release | 2002 | 0.446 | 0.63 | 0.87 | 7.59 | 1.00 / 1.00 | 5.85 | 0.80 / 0.77 | 8.93 |
| nonoise_s2001 | nonoise | 2001 | 0.285 | 0.63 | 0.83 | 7.70 | 1.00 / 1.00 | 5.55 | 0.79 / 0.77 | 8.81 |
| nonoise_s2003 | nonoise | 2003 | 0.040 | 0.60 | 0.93 | 7.81 | 1.00 / 1.00 | 7.47 | 0.95 / 0.97 | 6.43 |
| robust_s2001 | robust | 2001 | 0.307 | 0.64 | 0.80 | 8.04 | 1.00 / 1.00 | 5.64 | 0.80 / 0.78 | 8.97 |
| nonoise_s2004 | nonoise | 2004 | 0.273 | 0.65 | 0.85 | 8.15 | 1.00 / 1.00 | 6.79 | 0.80 / 0.80 | 9.02 |
| robust_s2002 | robust | 2002 | 0.881 | 0.64 | 0.80 | 8.91 | 1.00 / 1.00 | 6.85 | 0.79 / 0.74 | 10.47 |
| release_s2004 | release | 2004 | 0.925 | 0.60 | 0.67 | 9.81 | 1.00 / 1.00 | 9.52 | 0.78 / 0.73 | 11.63 |
| rel4 | released (one PPO run) | - | 0.025 | 0.95 | 0.99 | 10.30 | 1.00 / 1.00 | 13.70 | 1.00 / 0.99 | 9.76 |
| rel3 | released (one PPO run) | - | 0.021 | 0.97 | 0.97 | 10.69 | 1.00 / 1.00 | 13.87 | 1.00 / 1.00 | 9.20 |
| rel2 | released (one PPO run) | - | 0.018 | 0.87 | 0.99 | 11.65 | 1.00 / 1.00 | 14.50 | 1.00 / 1.00 | 9.18 |
| rel1 | released (one PPO run) | - | 0.010 | 0.49 | 0.69 | 13.50 | 0.89 / 1.00 | 15.61 | 0.87 / 1.00 | 10.58 |
Order by mean rank:
- Chrono: robust_s2003 > robust_s2004 > release_s2001 > nonoise_s2002 > release_s2003 > release_s2002 > nonoise_s2001 > nonoise_s2003 > robust_s2001 > nonoise_s2004 > robust_s2002 > release_s2004 > rel4 > rel3 > rel2 > rel1
- Nominal ROM: release_s2003 > nonoise_s2001 > robust_s2001 > release_s2001 > release_s2002 > robust_s2004 > nonoise_s2002 > nonoise_s2004 > robust_s2002 > nonoise_s2003 > robust_s2003 > release_s2004 > rel4 > rel3 > rel2 > rel1
- Primary TR: robust_s2004 > robust_s2003 > nonoise_s2003 > release_s2001 > release_s2003 > nonoise_s2002 > nonoise_s2001 > release_s2002 > robust_s2001 > nonoise_s2004 > rel2 > rel3 > rel4 > robust_s2002 > rel1 > release_s2004
Gap magnitude: Chrono time / nominal ROM time on goals both reach = 1.29 to 3.47 (median 1.45) across policies.
## Conditions and perturbation magnitudes
Units: action noise sd in normalized units (throttle range 0.2-0.5, steering range -1 to 1), per 0.1 s step; tau0 in N m; omega0 in rad/s;
steer_rate = largest change of normalized steering per 0.1 s step (Chrono loop: 0.1); delay and hold in 0.1 s steps; wheelbase in m.
| Condition | Role | Perturbation (one draw per trial; noise each step) | Trials per goal |
|---|---|---|---:|
| nominal | baseline | none (training ROM) | 1 |
| tr_rate | primary TR | tau0 ~ N(1000, 50); omega0 ~ N(2600, 100); sig_thr = 0.05; sig_steer = 0.05; steer_rate ~ LogU(0.05, 0.5) | 20 |
| paper_tr | secondary (noise-only TR) | tau0 ~ N(1000, 50); omega0 ~ N(2600, 100); sig_thr = 0.05; sig_steer = 0.05 | 20 |
| tr_rate_half | secondary | tau0 ~ N(1000, 50); omega0 ~ N(2600, 100); sig_thr = 0.05; sig_steer = 0.05; steer_rate ~ LogU(0.025, 0.25) | 20 |
| tr_rate_double | secondary | tau0 ~ N(1000, 50); omega0 ~ N(2600, 100); sig_thr = 0.05; sig_steer = 0.05; steer_rate ~ LogU(0.1, 1) | 20 |
| tr_rate_uniform | secondary | tau0 ~ N(1000, 50); omega0 ~ N(2600, 100); sig_thr = 0.05; sig_steer = 0.05; steer_rate ~ U(0.05, 0.5) | 20 |
| tr_delay | secondary | tau0 ~ N(1000, 50); omega0 ~ N(2600, 100); sig_thr = 0.05; sig_steer = 0.05; delay in {0, 1, 2, 3} | 20 |
| det_rate_0.1 | secondary | steer_rate = 0.1 | 1 |
| det_calibrated | secondary | wheelbase = 3.378; steer_rate = 0.1; hold = 10 | 1 |
## Post hoc summaries (not part of the fixed primary test, no new simulation; `posthoc_confirm.py`)
### primary pool
| Condition | tau(primary TR) - tau(condition) (95 % range) | Chrono-resolved pairs: same direction | opposite direction | unresolved |
|---|---|---:|---:|---:|
| nominal | +0.52 (+0.33 to +0.73) | 26/48 | 14 | 8 |
| tr_rate | - | 45/48 | 3 | 0 |
| paper_tr | +0.58 (+0.42 to +0.79) | 26/48 | 18 | 4 |
| tr_rate_half | +0.06 (-0.06 to +0.12) | 45/48 | 3 | 0 |
| tr_rate_double | -0.03 (-0.12 to +0.09) | 44/48 | 2 | 2 |
| tr_rate_uniform | -0.03 (-0.09 to +0.06) | 44/48 | 3 | 1 |
| tr_delay | +0.03 (-0.06 to +0.12) | 45/48 | 3 | 0 |
| det_rate_0.1 | +0.06 (-0.09 to +0.15) | 43/48 | 3 | 2 |
| det_calibrated | -0.09 (-0.21 to +0.15) | 43/48 | 0 | 5 |
Steering chatter (mean step change in nominal ROM) versus Chrono success, Spearman across policies: ID -0.03 (p = 0.928), OOD -0.76 (p = 0.004), all -0.73 (p = 0.008).
### secondary pool
| Condition | tau(primary TR) - tau(condition) (95 % range) | Chrono-resolved pairs: same direction | opposite direction | unresolved |
|---|---|---:|---:|---:|
| nominal | +0.15 (-0.03 to +0.25) | 80/102 | 14 | 8 |
| tr_rate | - | 79/102 | 10 | 13 |
| paper_tr | +0.20 (+0.02 to +0.28) | 80/102 | 18 | 4 |
| tr_rate_half | +0.57 (+0.37 to +0.58) | 55/102 | 43 | 4 |
| tr_rate_double | -0.17 (-0.33 to -0.12) | 98/102 | 2 | 2 |
| tr_rate_uniform | -0.12 (-0.28 to -0.08) | 97/102 | 3 | 2 |
| tr_delay | +0.18 (-0.05 to +0.28) | 75/102 | 7 | 20 |
| det_rate_0.1 | +0.55 (+0.37 to +0.60) | 51/102 | 43 | 8 |
| det_calibrated | -0.17 (-0.33 to -0.07) | 97/102 | 0 | 5 |
Steering chatter (mean step change in nominal ROM) versus Chrono success, Spearman across policies: ID -0.23 (p = 0.385), OOD -0.58 (p = 0.018), all -0.52 (p = 0.038).
## Notes
- Hidden-failure screening (Chrono success < 0.5) is not evaluable on set B: no policy had Chrono success below 0.5 (lowest: rel1 at 0.59, release_s2004 at 0.63).
- `report_confirm.py` and `posthoc_confirm.py` only format or summarize the outputs of `analyze_confirm.py`; they run no simulation.
|