henhao
Sign in
ZoologyJournal of Experimental Biology
Peer-reviewed journal article

Unsupervised learning reveals rapid gait adaption after leg loss and regrowth in spiders

Suzanne Amador KaneiD, Brooke L. QuinniD, Xuanyi Kris Wu, Sarah Y. Xi, Michael F. Ochs, S. Tonia HsiehiD

Abstract

Many arthropods and some vertebrates can voluntarily lose (autotomize) limbs during antagonistic encounters, and some can regenerate functional replacements. Spiders in particular frequently autotomize one or more legs. In this study, we investigated the time course of locomotor recovery after leg loss and regeneration in juvenile tarantulas (Arachnida: Araneae) with no prior experience of autotomy. We recorded high-speed video of spiders running with all legs intact, then immediately after, and again one day after, they autotomized two legs. The legs were allowed to regenerate, and the same sequence of experiments repeated. Running performance, posture, and path tortuosity were measured from video tracking. Spiders were found to resume their pre-autotomy speed and stride frequency after leg regeneration and in [≤]1 day after both autotomies; furthermore, path tortuosity was unaffected by these treatments. They adjusted their posture to compensate for missing legs, spreading their remaining legs and running with their bodies rotated 11-15 deg from their velocity. To analyze gaits, we applied unsupervised machine learning for the first time to measured kinematic data in combination with gait space metrics. Spiders were found to robustly adopt new gait patterns immediately after losing legs, with no evidence of learning. This novel clustering approach both demonstrated concordance with previously-hypothesized gaits and revealed transitions between and variations within these patterns. More generally, clustering in gait space enables the identification of patterns of leg motions in large datasets that correspond to either known gaits or undiscovered behaviors.

PDFDOI
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
8
FWCI
1.94
Field percentile
top 15%
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (24 Jan 2025) · CC BY 4.0

Gait-space clustering shows juvenile tarantulas losing two same-tetrapod legs immediately adopt variable tripod-like gaits, regaining pre-autotomy speed within a day without evidence of learning.

Leg autotomy is common in spiders, but prior arthropod studies disagree on whether limb loss slows running, rarely follow recovery over time, and had not tested how regenerated legs or repeated loss affect spider locomotion. This work addresses those gaps and adds hierarchical clustering in a toroidal gait-distance space so leg-phase patterns can be grouped without assuming predefined gaits.

Key results

  1. ShownLosing two legs did not slow running beyond the effect of cold anesthesia alone. Speed was back to normal by the day after each autotomy and matched pre-autotomy levels once the legs had regenerated.

    Kruskal-Wallis with Dunnett correction, N = 5 spiders: no speed difference C1 vs C2 (four of five regrown spiders filmed); day-of-autotomy drops in speed and stride frequency did not differ significantly from cooling-only controls (CC, N = 6) and were absent the next day; stride length unchanged; 43562 frames analyzed (96-173 strides per treatment). · Experimental procedure; Results; Running performance

  2. ShownSpiders missing two legs kept straight paths but changed posture: they ran with the body turned away from the direction of travel and spread the remaining legs to near-even angular spacing, with little change in how far the tarsi reached.

    Tortuosity did not differ among treatments (135 values in [0.896, 0.993], one outlier at 0.82); yaw near zero in C1 and CC vs 11-15 deg after autotomy; measured adjacent-leg angles matched uniform spacing (45 deg for 8 legs, 60 deg for 6); only significant range-of-motion change was a 1.3× area increase after the first autotomy; fore-aft and lateral extension unchanged. · Stability measures; Running performance

  3. ShownWard hierarchical clustering on normalized gait-space distances recovered model gaits from noisy simulations. It split the real data into three intact clusters near the alternating tetrapod and three autotomy clusters: two resembling the modified tripod and one resembling the ablated tetrapod.

    Simulated von Mises-perturbed ablated tetrapod and modified tripod samples were assigned correctly in 95% of cases by clustering vs 92% by nearest model gait; intact clusters ALT2, ALT1, ALT3 lay 11.5%, 14.6%, 20.4% of maximal distance from the tetrapod model; 62% of intact data fell in ALT1; autotomy data were dominated by MT1 and MT2. · Gait analysis; Gait analysis

  4. SuggestedA second autotomy showed no sign of error-based learning. Performance and gait use after the second leg loss resembled the first, which fits robust rather than adaptive neural control.

    Wilcoxon rank sum tests of 1AUT0 vs 2AUT0 grouped by specimen found no differences in speed, tortuosity or yaw; the share of data in MT clusters rose only 2.2% between first and second autotomy and fell 7.3% from day of to day after; per-cluster speeds similar for 1AUT0 and 2AUT0; speed correlated with time since treatment in only 1 of 24 trials. · Results; Running performance; Gait analysis; Discussion

  5. ShownMeasured gaits only roughly matched rigid alternation of leg groups. The cluster closest to the ablated tetrapod retimed its legs to avoid long two-leg support, and lower static stability in some clusters did not reduce speed.

    ABT1 support counts peaked at 3 and 4 tarsi rather than the predicted 2 and 4, with two or fewer tarsi grounded <27% of the time vs <2% for intact data; MT2 and ABT1 had lower SSM than MT1 yet similar speed; SSM fell significantly in all but one autotomy treatment; no aerial phase observed. · Running performance; Gait analysis; Discussion

Methods, data and reproducibility

Dorsal 500 frame/s video of juvenile tarantulas running intact, after two successive L4+R1 autotomies (same day and next day) and after regrowth; tracked body and tarsi yielded speed, tortuosity, yaw and stability, and per-frame interleg phase differences were clustered by Ward linkage on gait-space distances and compared with model gaits.

  • Captive-bred juvenile Guatemalan tiger rump tarantulas (2nd-3rd instar, 3-5 months old, autotomy-naive) given two successive L4+R1 autotomies with leg regrowth between · Davus pentaloris (Theraphosidae) · n = N = 5 spiders (regrowth trials for four); 45 purchased
  • Cooling-only control spiders filmed before and after cold anesthesia without leg loss · Davus pentaloris (Theraphosidae) · n = N = 6
  • High-speed dorsal running videos (up to 5 trials per specimen per treatment) tracked for center of mass, body axis and tarsal tips · Davus pentaloris (Theraphosidae) · n = 43562 frames; 96-173 strides per treatment · Supplementary Dataset 1
  • Synthetic interleg phase-difference data for alternating tetrapod, ablated tetrapod and modified tripod gaits with von Mises noise · n = 1000 samples per model gait
Methods & tools
MATLAB v2023b, Custom MATLAB automated tracking code (Appendix S1), DLTdv (Direct Linear Transformation data viewer), CircSTAT circular statistics toolbox, Photron FASTCAM SA-3 high-speed camera, Hierarchical agglomerative clustering with Ward's linkage, Gait-space distance metric on a hypertorus, Von Mises noise simulation (circ_vmrnd), Kruskal-Wallis test with Dunnett's post-hoc correction, Wilcoxon rank sum test, Bonferroni-Holm correction, Bootstrap confidence intervals (bootci, 1000 samples)
Reagents & environment
MATLAB v2023b (toolboxes: machine vision, curve fitting, statistics/machine learning); CircSTAT for circular statistics; DLTdv for manual checking and correction of tracks.
Reproducibility gaps
Data and code are stated to be provided only as Supplementary Dataset 1; no public repository URL or accession appears in the text, and custom tracking code details are deferred to Appendix S1.
Limitations (5)
  • Static stability margins were interpreted with caution because running tarantulas need not move quasi-statically and may stay stable through dynamic mechanisms or surface adhesion even when SSM < 0. · stated by the authors
  • Phase data by themselves cannot detect leg dragging, vertical body motion or dynamically stabilizing movements, so the authors advise pairing clustering with kinematics and annotated video; extension to 3D tracking is left to future work. · stated by the authors
  • Cooling-only controls were necessarily filmed first, on smaller and younger spiders, so they needed shorter anesthesia than the larger, older spiders that were cooled before autotomy. · stated by the authors
  • Only five spiders of one captive-bred species underwent autotomy (four contributing post-regrowth runs), so the null findings on speed loss and learning have limited statistical power and generality. · noted in this summary
  • The number of gait clusters was set by inspecting dendrogram branch heights and kinematic similarity rather than a formal selection criterion, so partitions such as MT1 vs MT2 may be analyst-dependent. · noted in this summary

Who should read:Read this if you study arthropod locomotion after limb loss or regeneration, or want an unsupervised clustering method for classifying interleg phase patterns in multi-legged gait data.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISPullar '18Song '20Owaki '26Hardesty '24Danner '23Pratt '24
64% semantic matchAnimal Behavior & Cognition
Kiri F. Pullar, Michael G. Paulin·bioRxiv·Aug 22, 2018