Summary
Researchers at Institute of Science Tokyo, Carnegie Mellon University and partner institutions have introduced InterPet4D, a dataset for modeling human–dog interaction, and InterPetMoGen, an AI system trained to generate dog motion from human body and hand movement plus audio. The dataset contains 6.83 million synchronized frames from 161 sessions involving 23 people and 13 dogs across 11 breeds, captured with 12 external cameras, an egocentric camera, audio and reconstructed 3D motion. According to the team, its full model reduced kinetic FID to 11.21 from a 21.22 transformer baseline and scored higher in a 12-person user study. The system currently creates fixed 10-second clips and does not model contact forces.
Hana’s perspective
I see the dataset as the more important contribution. Character animation systems often look convincing in isolation but fail when movement must respond to another body, a gesture, a voice and physical proximity at once. InterPet4D gives researchers a shared benchmark for that coupled problem, while the reported gains suggest multimodal conditioning can improve responsiveness as well as visual plausibility. The limits matter: 13 dogs are not the full range of canine behavior, contact forces are absent, and output is restricted to 10 seconds. Even so, this work points toward virtual animals that react to performers rather than merely replaying generic motion.
Source: Mirage News, “AI Captures Human-Dog Interactions to Generate Pet Motions,” https://www.miragenews.com/ai-captures-human-dog-interactions-to-generate-1756160/
Are you a writer with a great story? We'd love to hear from you: APPLY HERE