Regime Problem¶

Experiment 43's own addendum closed with seven independent, real-data-tested candidate fixes for the spike/regime labeling gap, all rejected -- and a specific conclusion: the signal being analyzed (one joint's leader-follower offset) probably doesn't contain the information needed to solve this with a cleverer function of itself alone. This experiment tests a genuinely different, physically-grounded idea instead of another function of the same signal: real mechanical backlash/deadband (a well-documented phenomenon in geared robot joints -- Lima, Machado & Crisóstomo, "Experimental backlash study in mechanical manipulators", Robotica 29(2):211-219, 2011, DOI:10.1017/S0263574710000056, checked via WebFetch before citing) as a confound that could be masquerading as some of the mislabeled spikes.

Step 1. Does the deadband signature exist in this real data?¶

Real robot joints briefly resist motion when the commanded direction reverses, until static friction is overcome -- not the paper's own pseudo-phase-plane+wavelet method, a much simpler classical signature check: real velocity right after a commanded reversal, against the joint's own baseline.

sign_change = np.sign(cmd_vel[1:]) != np.sign(cmd_vel[:-1])
post_rev_actual_vel = np.abs(act_vel[idx+1])  # right after each reversal
post_rev_actual_vel.mean() / np.abs(act_vel).mean()  # ratio vs. baseline

joint 0: ratio=0.313   joint 1: ratio=0.393   joint 2: ratio=0.157
joint 3: ratio=0.680   joint 4: ratio=0.347   joint 5: ratio=2.313

Real signature, on real data: 5 of 6 joints show real velocity dropping to 16-68% of baseline immediately after a commanded reversal. Joint 5 (the gripper -- an open/close mechanism, not a geared rotary joint) shows the opposite pattern, physically sensible, not discarded as noise.

Step 2. Do the flagged spikes correlate with deadband, across the whole dataset?¶

deadband_gate.py's deadband_mask flags frames plausibly inside a real deadband event. Checked whether classify_segments' spike-labeled points on Experiment 43's own signal (stable-frame joint-2 offset) are enriched inside those windows, across all 50 real episodes, not just the 2 hand-inspected ones:

50 episodes: total_points=4871, total_spike=394
spike points in deadband: 239 (60.7% of all spikes)
deadband base rate (all points): 34.6%
enrichment ratio: 1.75x

A real, generalized, 1.75x enrichment -- the first result in this whole line of investigation (Experiment 43's seven rejected fixes, plus the cross-channel-correlation experiment above) that holds up past the two hand-inspected episodes on first try.

Step 3. Does gating deadband points out actually fix the known mislabeled cases?¶

Experiment 43's addendum had already identified two specific spike-labeled runs confirmed (via net-displacement) to be real, sustained transitions: episode 0's run[90:95] and episode 22's run[24:29]. Re-ran classify_segments with deadband points removed from the signal first:

--- episode 22: 50/122 points gated out ---
  {'label': 'regime', 'start': 24, 'end': 28, 'median': -4.724}   <- WAS 'spike', NOW CORRECT
--- episode 0: 38/169 points gated out ---
  {'label': 'spike', 'start': 90, 'end': 94, 'median': -4.244}    <- STILL 'spike', unresolved

1 of 2 known cases fixed -- the first intervention in this entire investigation that moved a known-mislabeled case in the right direction at all. Not smoothed over: the other known case stayed wrong.

Step 4. Does it generalize, or was this two more hand-picked episodes?¶

Reused the same net-displacement ground-truth check from Experiment 43's own addendum (does a run's label match whether the signal settles at a new level or returns to baseline), across all 50 episodes, before vs. after deadband gating:

BEFORE gating: 135/207 runs match net-displacement ground truth (65.2%)
AFTER  gating: 57/92 runs match net-displacement ground truth (62.0%)

A small net regression (65.2% -> 62.0%), the same pattern every other candidate fix in Experiment 43 showed: real improvement on the 1-2 inspected cases, no improvement (here, slightly negative) once checked against the full dataset. Gating also structurally reduces the number of runs with valid clean neighbors on both sides (207 -> 92), which is itself part of why the aggregate comparison is noisier here.

Honest conclusion¶

The deadband/backlash confound is real, physically grounded (real citation, real signature verified directly in this data), and genuinely explains a disproportionate share of the spikes classify_segments flags (1.75x enrichment, generalized across all 50 episodes) -- a real, distinct finding from Experiment 43's own spike/regime duration problem, not a restatement of it. But using it as a fix for that specific problem does not work: it corrects one of the two known mislabeled cases while leaving the other wrong, and makes the aggregate metric slightly worse, not better. The two problems (deadband noise vs. genuine short-duration real transitions) coexist in the same signal but are not the same problem, and fixing awareness of one does not fix the other. Not pursued further as a fix for Experiment 43's open gap; the deadband/spike association itself is worth keeping as a separate, real, disclosed finding.

Reproducing this¶

scripts/robot_sensor_validation/deadband_gate.py (deadband_mask) -- reuses the same cached lerobot/svla_so101_pickplace dataset as Experiments 43 and 46, no new download. Real literature grounding this page cites is indexed in quantumrag's robotica_rilevamento_anomalie collection alongside Experiment 46's paper.