Real-time body pose non-verbal communication with a consistency-based reliability measure

A new study proposes recognizing communicative intent from 2D body pose alone, releasing a dataset of real full-body frames covering ten intents and targeting low-cost, real-time person-to-robot communication on embedded hardware [1]. The work, posted to arXiv on June 8, 2026, argues that body motion remains a reliable signal in conditions where face and speech cannot be captured, such as long-distance rescue missions [1][2]. Existing resources do not isolate this signal: affective corpora combine body, face, voice, and text, while skeleton action-recognition benchmarks label the action performed rather than the message conveyed [1][2]. The researchers release a dataset of real frames of full-body pose covering ten communicative intents and compare it against other real and synthetic datasets, including IPC, MotionLCM, VEO3.1, and Kimodo, which span a range of difficulty [1][2]. The study targets systems that can run on a robot's limited onboard hardware, benchmarking multiple models—from skeleton graph classifiers to joint motion-forecasting networks—and reporting performance metrics together with frame rate on an embedded GPU, the NVIDIA Orin Nano [1][2]. A key contribution is the demonstration that a model's own autoregressive self-consistency can serve as an unsupervised reliability signal [1][2]. The authors provide a short proof that bounds the probability that a self-consistent prediction is correct, show that this probability grows with the number of consistent steps, and identify the conditions under which a confident prediction can still be false, benchmarked against industry-standard metrics [1][2]. The approach aligns with the broader scientific method, which relies on hypothesis testing and experimental validation to refine models [5]. In practical terms, reliable non-verbal communication could reduce errors in high-stakes environments, much as patient safety initiatives have sought to systematically prevent harm through better reporting and analysis [3]. The work also touches on selection and assessment processes familiar from structured job interviews, where evaluators must interpret candidate signals under constrained conditions [4].