Companion code and data for the PGOB manuscript (Parameter Generation by Observed-Behaviour), sole author Chenxi He, Cavendish Laboratory, University of Cambridge. PGOB obtains a working simulator parameter set non-invasively, from ordinary tracking video in place of physiological measurement; the inference engine (SPSA, CMA-ES, or amortized simulation-based inference) is interchangeable. Primary result: non-invasive, behaviour-equivalent parameter generation from a video-only uninformative start. Trajectory-only generation is demonstrated on three published simulators (BAAIWorm, modWorm, FlyGym); flybody is a boundary case (MoCap-locked walk-imitation, parameter signal in loss space, trajectory arms within <1.5%). Against real video the generated parameters are non-inferior to the simulator's own invasively-calibrated default (they match, not beat, it). The headline claim is behaviour-equivalent parameters, not the animal's true physiological values. This v1.2 archive expands the recovered checkpoints: per-mode theta vectors, BAAIWorm SOTA/full5 and per-ablation checkpoints, the amortized neural-posterior models and samples, recovered behavioural trajectories, the genuine result files (three-distance, per-feature within-variance, cross-species), the real reference tensors (OWMD-N2, Janelia), the GAN-as-loss negative-control checkpoints (SI only), and the final manuscript PDFs (NCS article, comprehensive, supplementary). Code: https://github.com/ChenxiHeCam/Parameter-Generation-by-Observed-Behavior . Data public under original licences; code MIT.
Publication Date: 2026-06-20