For decades, the dominant explanation for crowd disasters invoked panic. Early theories suggested that people in emergencies lose rational control and behave selfishly or chaotically, and simulation models were built to propagate emotional states through virtual crowds. But empirical evidence has steadily dismantled this picture. Eyewitness accounts, video analyses, interviews and sociological investigations of real disasters consistently show that people in extreme crowding remain calm, cooperative and often altruistic. Fatal outcomes instead arise from collective mechanical instabilities, when local pressures, force chains and density waves propagate through a compressed mass of bodies, causing loss of balance, fainting and asphyxiation. The crowd, in this reframing, is a complex dynamical system governed by biomechanics, not by emotion.

The empirical record of dense-crowd behaviour is remarkably rich. Observations during the 2006 Hajj in Mecca documented stick-slip instabilities and turbulent-like flows, sometimes described as earthquake-like crowd turbulence, in which sudden large-scale ruptures sweep through packed pedestrians. Propagating density waves were famously recorded at an Oasis concert in the United Kingdom in 2005, where waves emerged spontaneously in the standing audience. Transversal oscillating rushes have been measured experimentally in competitive crowds pressing toward bottlenecks, and collective chiral rotations, in which an entire crowd slowly turns as a coherent body, have been documented at the St. Fermín festival in Pamplona and at the Love Parade in Duisburg. These patterns arise without clear triggers and without intentional coordination, and no existing model could reproduce them within a single interpretable framework.

The reason, the authors argue, lies in a simplification that nearly all existing models share: the representation of pedestrians as passive two-dimensional disks. Classical microscopic models, built on social forces, collision avoidance or trajectory optimisation, work well at low and intermediate densities, where balance is maintained effortlessly through normal locomotion. But at densities of four to five pedestrians per square metre and above, contact becomes continuous and the dynamics change fundamentally. Contact forces challenge postural control, standing balance must be actively maintained, and the upper body and legs become constrained in different ways. A disk cannot topple, cannot step, and cannot fall, so a crowd of disks cannot capture the biomechanics that dominates motion at extreme density.

The new model, deliberately minimalist, gives each simulated pedestrian two coupled subsystems: an upper body and a set of legs, each represented by a position in the horizontal plane. The upper body corresponds roughly to the ground-projected centre of mass, while the legs correspond to the centre of the support area beneath the feet. Two antagonistic mechanisms couple them. When the upper body drifts away from the legs, an unintentional unbalancing force, effectively the linearised effect of gravity, amplifies the displacement, just as a leaning body tends to fall. In response, an intentional rebalancing force drives the legs to move back underneath the upper body, restoring the upright posture. Each pedestrian is thus a mechanical analogue of an inverted pendulum, a conceptual model long used in individual postural-control research, but here embedded in a dense interacting crowd.

Interactions between pedestrians are handled through purely repulsive, short-range potentials acting independently at the two levels, with a larger repulsion distance at the upper-body level than at the leg level, reflecting the fact that shoulders are wider than feet. This asymmetry makes the legs more mobile and better able to adjust under compression. A damping term represents each pedestrian’s active resistance to being pushed. Crucially, the collective dynamics turn out to be governed mainly by the balance and unbalance rates and speeds rather than by the precise form of the contact potentials, which work equally well whether exponential or algebraic. The organising principle, in other words, is internal balance control, not the fine details of contact modelling.

When the researchers simulated 196 pedestrians in a small periodic domain, the model spontaneously produced collective patterns that had eluded previous modelling attempts. With intermediate unbalancing rates and speeds, coherent diagonal density waves emerged and propagated across the crowd at roughly four to five metres per second, matching the wave speeds observed at the Oasis concert and in controlled experiments on impulse propagation through standing crowds. With a higher unbalancing rate and a lower speed, the system instead organised into a global chiral rotation, with every pedestrian tracing a synchronised circular trajectory at a period of about twelve seconds, strikingly close to the roughly eighteen-second rotations documented in Pamplona. The rotational period proved tunable through the model parameters, with the ratio of unbalancing to balancing rates controlling the timescale.

A systematic exploration of the parameter space revealed a well-structured phase diagram with four regimes: a crystallised state with low energy and low correlation, a density-wave state with intermediate energy and high coherence, a chiral-oscillation state with low energy and high coherence, and a disordered state in which body and leg behaviours decouple. The authors tested robustness by varying system size, adding positional noise to initial conditions and drawing individual parameters from random distributions; the phase diagram remained essentially unchanged. Collective dynamics also persisted without leg-to-leg interactions, although synchronisation became harder, and were insensitive to the damping rate provided it remained positive. The emergent patterns, in short, are intrinsic features of the coupled balance dynamics rather than artefacts of fine-tuning.

Mechanically, the crowd can be viewed as a spatially extended network of coupled inverted pendulums, or equivalently as mass-spring-damper units exchanging energy and perturbations through short-range exclusion forces. This interpretation connects the model to the theory of coupled oscillatory systems, multi-layer dynamics and Kuramoto-type synchronisation, offering a conceptual bridge between individual postural control and crowd-scale order. The internal two-layer structure introduces degrees of freedom that naturally favour oscillatory responses and feedback loops, creating precisely the conditions under which macroscopic modes can emerge from purely local interactions, without any global rule or intentional coordination.

The authors are careful to position the work as explanatory rather than predictive. The full human body dynamics at stake in a real crush are far more complex, and future extensions must address the singularity that arises when body and legs coincide, replace the propulsive unbalancing force with a conservative gravitational-like potential, incorporate state-dependent parameters, heterogeneous populations, friction and falling mechanisms. The role of the arms remains an open question: evidence from crowd disasters shows suffocation by thoracic compression is a leading cause of death, and eyewitnesses describe struggling even to protect their chest, yet the model suggests active arm pushing is not required to reproduce observed collective motion. Calibration against video data of large events and controlled motion-capture experiments will determine whether this two-level paradigm, which gives crowd particles legs, can mature into a quantitative tool for forecasting and mitigating hazards in the world’s most dangerous gatherings.

Subject of Research: Biomechanical balance recovery as a driver of emergent collective dynamics in ultra-dense pedestrian crowds

Article Title: Giving particles legs: How balance recovery may drive emergent collective dynamics in ultra-dense crowds

Article References: Chatagnon, T., Chraibi, M., Pettré, J., Seyfried, A., & Tordeux, A. (2026). Giving particles legs: How balance recovery may drive emergent collective dynamics in ultra-dense crowds. PLOS Complex Systems, 3(10), e0000110. https://doi.org/10.1371/journal.pcsy.0000110

Image Credits: AI Generated

DOI: 10.1371/journal.pcsy.0000110

Keywords: crowd dynamics, ultra-dense crowds, balance recovery, pedestrian model, inverted pendulum, density waves, chiral rotation, emergent behaviour, biomechanics, crowd safety, self-organisation, PLOS Complex Systems

The research, led by Andrew J. Gunnell, Owen Winship, Lukas Gabert and senior author Tommaso Lenzi at the University of Utah, brought eight above-knee amputees and eight non-amputee controls into a motion capture laboratory and subjected them to a battery of precisely controlled sideways tugs. Participants walked on an instrumented split-belt treadmill while a cable-driven perturbation system, built from the open-source BumpEm design and controlled through custom LabVIEW software, delivered sharp 300-millisecond pulls to a harness worn at the waist. The cables attached at approximately the height of each person’s center of mass, near the navel, and the motors were positioned so the pulls remained perfectly horizontal. That detail matters: a cable anchored too high would twist the trunk, and one anchored too low would act on the legs rather than the whole body. Between perturbations, the cables maintained a gentle six-newton pre-tension so no slack could dull the force.

The experimental matrix was exhaustive. Each participant experienced twelve distinct perturbation conditions: pulls delivered at three different moments in the gait cycle (heel strike, late swing, and late single support), from two directions (toward the prosthesis side or the sound side for amputees, left or right for controls), and at two force levels, roughly 7.5 and 15 percent of body weight. Each condition was applied once per trial across five randomized trials, totaling 60 perturbations per session, with random three-to-ten-second delays to prevent anticipation. Walking speed was standardized at 0.9 meters per second, though two amputees needed to slow to 0.7 meters per second. Twelve Vicon cameras tracked 65 reflective markers at 200 hertz while force plates under each foot recorded ground reaction forces at 1000 hertz, allowing the team to reconstruct whole-body biomechanics in extraordinary detail.

To quantify instability, the researchers turned to integrated whole-body angular momentum, or iWBAM, a validated metric that captures the body’s cumulative rotational dynamics around its center of mass. They computed iWBAM for every steady-state step during a three-minute baseline walk and for the first step after each perturbation, then normalized each perturbed step against the person’s own baseline variability. The Euclidean distance between a perturbed step and the steady-state mean, expressed in standard deviations, yielded a continuous instability score, binned into four levels: no response, low, medium, and high. A fall was defined as grabbing the handrails with more than five percent of body weight or stepping off the treadmill belt. Recovery was declared only when whole-body angular momentum re-entered the person’s steady-state variability band, within four standard deviations, and stayed there for at least 400 milliseconds.

The headline result is stark. Non-amputee participants never fell, not once, across every condition. Above-knee amputees fell 45 times out of 471 perturbation trials, a fall rate of 9.6 percent. Every single fall came from a high-force pull, and 42 of the 45 occurred during the late single support timing, when the body is balanced on one leg with the other limb swinging forward. Within that most dangerous condition, statistical modeling showed that direction mattered enormously: amputees were significantly more likely to fall when pulled toward their prosthesis side than toward their sound side. Sixteen falls occurred on sound-side pulls and 29 on prosthesis-side pulls, a pattern that ultimately traces back to the mechanical limitations of prosthetic knees and the weakened residual hip.

The instability analysis added a subtle twist. When averaged across all conditions, amputees did not show uniformly worse stability than controls. Instead, the group difference emerged through interactions: high-force pulls and late single support timings hit amputees disproportionately hard, with a significant group-by-timing-by-magnitude interaction confirming that this population is especially sensitive when difficult factors combine. Surprisingly, during late single support, pulls toward the sound side actually produced greater measured instability than prosthesis-side pulls, yet prosthesis-side pulls caused more falls. The resolution to this apparent paradox lies in the recovery strategies each group could deploy, and in the strategies that were simply off the table.

During early and mid-cycle perturbations, both groups relied on similar lateral stepping strategies and amputees fell only twice. The trouble came during late single support. Non-amputees facing high-force pulls shifted away from ankle-and-hip joint strategies toward cross-behind steps and even jumps, using their two biological legs to rapidly reposition their base of support. Amputees pulled toward the prosthesis side had no such freedom. Jumping was essentially impossible because a passive prosthetic knee and ankle cannot generate the positive power needed to leave the ground, and joint strategies demand that the weakened residual hip supply all the corrective torque alone. That left only crossover and cross-behind steps, both of which require rapid weight shifting between limbs and precise interlimb coordination. When amputees attempted crossover steps under high force, they fell in 67 percent of prosthesis-side trials and 100 percent of sound-side trials; cross-behind attempts fared only marginally better at 62 percent.

Sound-side pulls told a different story. Because crossing the prosthesis behind the sound limb was effectively unavailable, as microprocessor knees offer little resistance early in swing and may fail to transition correctly into the high-resistance stance phase, amputees leaned heavily on jumping, attempting it for roughly 40 percent of high-force sound-side pulls compared with just 17 percent for non-amputees. Jumping worked remarkably well as a fall-prevention tactic, with a fall rate of only 5.8 percent, which helps explain why sound-side pulls produced more measured instability but fewer falls. The catch is metabolic and orthopedic: jumping places extreme loads on the sound-side knee, and above-knee amputees already face elevated osteoarthritis risk. The researchers suggest that this habitual reliance on a high-cost strategy may quietly accelerate joint wear over years of daily ambulation.

The findings carry direct implications for rehabilitation and device design. The authors argue that physical therapy should target the perturbation conditions most likely to cause falls, particularly medial stepping maneuvers, and that hip strengthening programs could make joint strategies viable at higher perturbation forces, reducing the need for desperate jumps. On the engineering side, they propose that microprocessor knee algorithms could automatically increase extension torque capacity when they detect sudden lateral pelvic acceleration, preventing the prosthesis from buckling during a reactive step. Powered knee-and-ankle prostheses and frontal-plane hip exoskeletons, both active areas in Lenzi’s laboratory, could further expand the recovery toolkit by supplying the push-off power and hip torque that passive devices cannot deliver.

The study is not without limitations. Treadmill walking at a fixed, slightly sub-preferred speed, constrained arm movement, narrow belt width, and the presence of handrails all diverge from real-world conditions and may have shaped the observed strategies. The cohort consisted exclusively of highly active K3 and K4 ambulators using a variety of microprocessor knee models, so results may differ for less mobile users or specific devices. Still, as the first mediolateral waist-perturbation study conducted on above-knee amputees during treadmill walking, the work fills a critical gap and offers engineers and clinicians a condition-by-condition map of exactly where and why balance breaks down, a roadmap that could guide the next generation of fall-prevention technology for millions of prosthesis users worldwide.

Subject of Research: Balance recovery strategies of above-knee amputees after mediolateral walking perturbations

Article Title: Above-Knee Amputees’ Balance and Recovery Strategies After Mediolateral Perturbations from a Cable-Driven System

Article References: Gunnell, A. J., Winship, O., Williams, A. E., Murray, R., Stoddard, G. J., Gabert, L., & Lenzi, T. (2026). Above-Knee Amputees’ Balance and Recovery Strategies After Mediolateral Perturbations from a Cable-Driven System. Annals of Biomedical Engineering. https://doi.org/10.1007/s10439-026-04355-8

Image Credits: AI Generated

DOI: 10.1007/s10439-026-04355-8

Keywords: above-knee amputees, balance recovery, gait perturbation, prosthetic knee, whole-body angular momentum, fall risk, cable-driven system, biomechanics, microprocessor knee, treadmill walking, recovery strategies, assistive technology