A Mechanical Threshold and Divergence Law for Collective Transport: Closed-Form Theory, Rigid-Body Simulation, and Decentralized Reinforcement Learning
Year: 2026
Type: preprint
Abstract: We study collective transport, in which N simple robots jointly push a rigid body toward a goal. From elementary rigid-body dynamics and Coulomb friction we derive a closed-form mechanical threshold N * = µmg/F max below which the body cannot leave rest, and show that the time to reach a goal at distance d diverges as T (N) ∼ (N − N *) −1/2. A Newton–Euler simulation with stick–slip friction recovers α = 0.500 to within 0.1% in five physical regimes, with errors that decrease roughly in proportion to the time step. We then test whether the law survives decentralized, communication-free control with noisy local sensing, and derive a closed-form noise-coherence correction N * eff = N * e σ 2 /2 for controllers that do not compensate their sensing errors. In the simple setting, the central methodological finding is a control: an untrained identity policy (applied force direction equal to the sensed goal direction) already attains 100% success, α ≈ 0.49 and N 0 ≈ 29.8. Consequently, in that setting success and exponent recovery by a trained policy are not evidence of learning. With the goal direction randomized, 20 independently trained policies converge to this identity map (sensitivity slope 1.00 ± 0.00) and the correction matches the observed threshold within 0.5% at every noise level tested (σ step ∈ [0, 0.5] rad); with a fixed goal, trained policies instead partly memorize the goal and fall below the correction. Masked-input negative controls behave as controls only when the goal is randomized. Rotational dynamics do not alter the translational law. Finally, where learning has something to learn, i.e. in a harder task with large sensing error and an offset centre of mass, we train decentralized policies with MAPPO (residual-angle action, threshold curriculum, recurrent memory); they far exceed standard MARL baselines and slightly exceed tuned designed controllers at the training condition (success 0.813 ± 0.025 over 10 seeds versus 0.758 and 0.781), but are worse than those controllers at low noise, and the margin over the strongest learned baselines is not significant on fresh seeds. All results are simulation-only.
Source: Research Square
Authors: Milad Tahavor, Mohsen Rabdoost Motlagh
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236351430
