WoRD uses diffusion models to infer goals from partial trajectories
Researchers describe a goal-recognition method that runs a denoising chain per candidate goal and reports calibrated results on simulated and real movement data.
In plain words
Goal recognition means guessing what a person or robot is trying to achieve by watching how they move. The WoRD method learns how agents typically travel, then checks each possible goal against the observed path and combines the results into a probability score. The authors say it stays reliable when only the start of a path is visible.
What was announced
The paper's authors argue that existing approaches split into two camps: symbolic methods that ignore continuous motion, and trajectory-based methods that need an explicit dynamics model which is difficult to obtain and does not transfer well between settings. Neither, they write, learns how agents actually move toward each possible goal, which hurts accuracy when only an early, ambiguous stretch of a path is available. WoRD addresses this with a learned prior over trajectories. At every step it runs one warm-started denoising chain for each candidate goal, pulls the chains toward the observations using a closed-form log-likelihood gradient, and applies Bayes' rule to the samples to produce a posterior over goals.
Evaluation covered simulated robot scenarios, pedestrian data, and vehicle data, with four metrics chosen to probe different failure modes. According to the abstract, the identical inference procedure kept beliefs calibrated and multimodal through ambiguous path prefixes, while filter-based and classifier-based observers either collapsed onto a single outcome or became overconfident. The reported evidence is limited to the paper's abstract, so experimental details, baselines, and any code availability are not described here.
What it does not show
- Evidence comes only from the paper's abstract; full experimental setup, baselines, and results are not available in the supplied passages.
- All performance claims are the authors' own and have not been independently confirmed.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.