FC: fall-criterion correction and re-scoring (registered: CRITERIA_FC.json, 2026-11-11)
Why
Every runner had inherited from the project’s first runner a stop “pelvis tilt > 0.35 rad or pelvis height < 0.5 m”, counted as a fall, with no documented source.
On dev seeds 1300-1302 (lift motion + 50 N fore-aft push), the final controller hit the 0.35 rad stop 3/3 with both feet loaded and the DCM inside the support.
With the stop at 0.6 rad the same runs recovered 3/3 without a step (pelvis peak 0.37 rad). The stop is not neutral between strategies: an in-place hold may lean the trunk, while a step keeps it upright.
Criterion
- Fall (first occurrence of any of):
- a non-foot body touches the ground;
- pelvis height < 0.5 m;
- tilt > 1.0 rad (divergence stop only);
- a controller error.
- Recovered: no fall, and standing at the end (last 0.5 s: mean pelvis height > 0.6 m and mean horizontal speed < 0.1 m/s).
- Steps: counted until the fall.
Procedure
- Runs that ended on the old stop (
physical_bounds) were re-run with permissive stops (tilt 1.2 rad, height 0.4 m), suffix _fc.
- Scripts:
claude_brace/run_fcall.sh MODE=fc, the MT_FC hook in run_mt.sh.
- All other runs were re-scored from their logs.
claude_brace/fc_rescore.py writes claude_brace/runs_fc/. The registered analysis scripts were then run unchanged on that tree (FC_RUNS=...) and wrote *_results_fc.json.
- Factorial HFDB/HFABL (seeds 1090-1099) used the variants reported in the manuscript: B11, SME, R1i, DB, 65 re-runs.
Reproduction check (_rep, original stops)
- Every experiment reproduced exactly, except one LN2D run,
mt1_SmePI_100_0.3_i+50_s1261_ln2Ix_ORd20n0:
- stop time 6.525 vs 6.500 s; failure type and 9 steps identical;
- divergence began at 4.8 s, after 9 steps.
- This is reported; the outcome is the same.
Counts (manuscript experiments, registered seed ranges; fc_counts.json, fc_sens_paper.json)
- 4372 runs re-scored.
- 40 changed failed → recovered; 22 changed recovered → failed.
- The 22 are mostly end-state
not_standing (NZ1, NZ2, EST1).
- MOTCI: 5 of the OURS runs changed failed → recovered (lift + forward push).
| Experiment |
n |
Old recovered |
FC recovered |
FC held |
| BUD2 |
216 |
191 |
191 |
88 |
| BUDF |
240 |
180 |
181 |
93 |
| COST1 |
348 |
330 |
330 |
214 |
| EST1 |
380 |
316 |
315 |
212 |
| EST2 |
152 |
131 |
131 |
89 |
| HFDB (B11, SME, R1i, DB) |
320 |
248 |
249 |
22 |
| HFFINAL |
848 |
680 |
689 |
353 |
| HFLAT |
192 |
167 |
170 |
94 |
| HFLAT3 |
208 |
178 |
184 |
123 |
| LAT2 |
608 |
425 |
425 |
311 |
| LN2D |
204 |
180 |
181 |
97 |
| MOTCI |
144 |
129 |
132 |
101 |
| MOTCI2 |
192 |
149 |
152 |
95 |
| NZ1 |
240 |
188 |
184 |
116 |
| NZ2 |
80 |
75 |
71 |
53 |
Decisions
- All registered decisions are unchanged, except MOTCI M3 and M4, which were not met before and are now met:
- M3, held: 30 vs 23, discordant 7 vs 0, p = .0078;
- M4, recovered: 43 vs 42.
make_numbers.py FC=1 reports “FC changed decisions: []”: every asserted decision holds.
Factorial steps (seeds 1090-1099, 80 runs per variant, steps until the fall)
| Variant |
Old steps |
FC steps |
| B11 |
210 |
231 |
| hold layer alone (SME) |
186 |
198 |
| IMU estimate alone (R1i) |
219 |
244 |
| both (DB) |
141 |
159 |
The order is unchanged. Totals rose because failed runs now continue past 0.35 rad and step before falling.
Sensitivity (reported, not a decision)
- Tilt stop 0.6 / 1.0 / 1.2 rad × height 0.4 / 0.5 m: counts identical.
- End speed 0.05 m/s: counts change (MOTCI, MOTCI2, HFFINAL), but all decisions are unchanged.
- Tree:
runs_fc_v05, results *_fcv05.json.
- Controller touchdown timeouts:
- 381 runs end on one, 89 of them at tilt < 0.35.
- They count as failures for every model-based controller alike.
Other threshold audits (offline, robust_holdability/estcmp3_results_*.json)
- Detection threshold 15 / 25 / 35 N: the order of the estimators is unchanged.
- Ours is faster than MOB and EMO at every threshold.
- At 15 N, ours vs CIMU is n.s. (28.0 vs 28.3 ms, p = .09).
- Replay noise ×2: unchanged.
- Replay noise “severe”: ours 21.2 N vs CIMU 21.8 N RMS (p = .2, n.s.); MOB and EMO remain slower and less accurate.
- ±0.08 tolerance: about one 5 N step of the boundary force grid. This is stated as a fact, not as the original rationale.
Manuscript
manuscript_RAL/main.tex now uses numbers_fc.tex and figures/fig_requirement_fc.pdf (FC=1 python3 paper_figure/make_requirement.py).
- The original-criterion numbers stay in
numbers.tex and *_results.json.
Touchdown-timeout diagnosis (2026-11-11, exploratory)
What the stop is
- The “touchdown timeout” is raised in
ctrl_v2/autostep.py when the swing foot has not touched down 0.35 s after the planned landing.
How the runs looked at the stop
- In the paper experiments, 421 runs ended this way.
- At the stop:
- pelvis tilt, median 0.54 rad (IQR 0.37–0.71);
- horizontal speed, median 0.63 m/s.
- Only 37 runs (9 %) were “calm” at the stop: tilt < 0.35 rad and speed < 0.3 m/s.
Re-run with the timeout disabled
- Which runs: all 37 calm runs plus 40 random others (
random.seed(1)).
- Settings:
- opt-in
TD_TIMEOUT=5 (run_mt.sh MT_FC=td, suffix _td);
- permissive stops;
- scored with the FC criterion.
- Result: none of the 77 recovered.
- 69 fell physically (calm 35, moving 34);
- 7 ended on a tracker QP failure;
- 1 was not standing.
- The stop therefore marks a fall in progress, not a stand that was cut short.
- Results file:
td_diag_results.json.