AI generates realistic dog movements from human-dog interaction data

Researchers at the Institute of Science Tokyo, in collaboration with Carnegie Mellon University and partner institutions, have introduced InterPet4D, a comprehensive multimodal dataset designed to capture three-dimensional human-dog interactions over time. The findings, detailed in a recent arXiv preprint and scheduled for presentation at ECCV 2026, address a critical gap in artificial intelligence by focusing on the complex dynamics of human-animal engagement. Traditional motion capture struggles with close-range interactions where occlusion often hinders detailed reconstruction. To overcome this, the research team compiled 161 recording sessions featuring 23 human participants and 13 dogs across 11 distinct breeds. Using twelve synchronized third-person cameras alongside egocentric head-mounted units, the team captured multi-view video, spatial motion, aligned audio, and descriptive text. These interactions span four primary categories: petting, verbal commanding, distance calling, and free-form activities such as fetch and tug-of-war. This multimodal approach provides the temporal and contextual data necessary for systematic behavioral and computational analysis. Leveraging InterPet4D, the researchers developed InterPetMoGen, an artificial intelligence framework capable of generating realistic canine movements based on human gestures and vocal cues. The model processes complex physical actions into compact motion tokens and utilizes an autoregressive transformer architecture. By integrating a PetVAE autoencoder with modality-aware attention mechanisms, the system dynamically weighs input signals to produce contextually appropriate responses. Performance benchmarks demonstrate substantial improvements over existing methods. InterPetMoGen achieved a kinetic Fréchet Inception Distance score of 11.21, representing a 47.2 percent reduction compared to a standard Seq2Seq-Transformer baseline, indicating significantly higher fidelity to real-world motion. The framework also improved hand and body motion alignment metrics while increasing the diversity of generated sequences. In blinded user evaluations, the system scored 6.58 for naturalness and 6.63 for overall quality on a seven-point scale, far surpassing comparative models. The initiative establishes a standardized foundation for studying interspecies interaction, with immediate applications spanning computational biology, character animation, interactive virtual agents, and socially aware robotics. While the current iteration generates fixed-duration ten-second clips without simulating physical contact forces, the researchers intend to expand the framework to accommodate other animal species, model tactile interactions, and support longer temporal sequences. This work marks a pivotal step toward AI systems capable of perceiving and responding to the nuanced dynamics of human-pet relationships.