A model designed to forecast several plausible human movements reports lower errors on laboratory and outdoor benchmarks, alongside modest gains from observing nearby soccer players. FlowMoP, developed by Aadya Agrawal, Ho Kei Cheng and Alexander Schwing at the University of Illinois Urbana-Champaign, was posted to arXiv on November 6, 2026. It is an author-reported preprint; peer review has not been verified, and no independent replication is supplied.
The problem is that a person's recent movement can lead to several different futures. FlowMoP uses conditional flow matching to transform random noise into possible future motions, representing movement through both the hip's global trajectory and joint positions relative to that root. B-spline control points provide a compact, temporally smooth representation, while a dual-stream transformer processes observations and evolving predictions separately. Despite the abstract's description of joint generation, the detailed method predicts each person separately, optionally using neighbors' observed trajectories. Enforcing consistency among everyone's predicted futures remains future work.
For evaluation, FlowMoP generates 100 proposals using 10 Euler integration steps, then clusters them to select six representatives. Each error metric takes the lowest error among those six against the recorded future. This best-of-six procedure measures whether a close candidate is available; it does not show how accurately the system can select one forecast in deployment or whether its probabilities are calibrated.
On CMU-Mocap (UMPM), which uses one second of observations to predict two seconds of movement, FlowMoP reported two-second joint position error (JPE), root-relative pose error (APE) and final root displacement error (FDE) of 237.4, 128.6 and 159.3 millimeters. Corresponding errors were 250.4, 146.5 and 171.6 millimeters for EMPMP, and 275.0, 161.1 and 199.2 millimeters for the stochastic DuMMF model. EMPMP and DuMMF were recomputed for these comparisons.
On the outdoor 3DPW benchmark, using 0.8 seconds of observations, FlowMoP's errors at 1.6 seconds were 232.1, 146.8 and 150.5 millimeters, versus EMPMP's 235.4, 150.6 and 153.0 millimeters. These smaller differences lack confidence intervals, repeated-training variability estimates and significance tests, leaving their robustness uncertain.
WorldPose tests professional soccer scenes with 22 players and poses estimated from broadcast video. Using one second of observations to forecast three seconds, adding ten observed neighbors reduced FlowMoP's JPE from 1.87 to 1.82 meters and FDE from 1.79 to 1.75 meters. APE remained 0.10 meters at reported precision. This supports a modest benefit from social context in this evaluation, while video-derived reference poses introduce measurement uncertainty. Selected trajectory illustrations do not establish dataset-wide uncertainty calibration or causal effects of social interactions.
A CMU representation ablation also favored cubic splines over raw frames: JPE was 237.4 versus 245.8 millimeters. A single forward call took 6.61 ± 0.40 versus 7.86 ± 0.22 milliseconds across 50 runs at batch size 64. Those timings exclude complete sampling, integration and clustering. More broadly, agent windows share scenes and recordings, so they cannot automatically be counted as independent observations; independent sample counts are not supplied. The results support benchmark improvements, with deployment reliability and mutually consistent group forecasting still unresolved.
Cover illustration generated with AI; conceptual, not research evidence.