Researchers have found a way to stop animated humans from clipping through chairs and tables when they interact with objects in 3D scenes.
A new method called MAMHOI splits scene-aware human-object interaction generation into two separate steps instead of asking one model to handle everything at once. First, a scene-conditioned model figures out where and how an interaction is physically feasible within a given room, essentially mapping out usable affordances in the environment. A second model, conditioned on that affordance information, then generates the actual human and object motion for the interaction. The researchers built this split because training data pairing humans, objects, and full 3D scenes together is scarce, while human-scene movement data and human-object interaction data each exist separately in larger supply. Tests in complex indoor environments showed the approach cuts down on objects and people visually overlapping with scene geometry while keeping the interactions looking more natural.
This is a data problem as much as a modeling one. Instead of waiting for someone to collect an expensive triple-matched dataset, MAMHOI squeezes more value out of data that already exists by training the two halves separately and connecting them through a shared affordance interface. That approach could matter for anything that needs characters to convincingly use their surroundings, including game animation, robotics simulation, and training data for embodied AI.
It is still a research prototype measured on indoor benchmarks, not a system shipping in any game engine, so the real test is whether this factorization holds up outside a lab's curated room layouts.