← ArchivePaper2022
The Wanderings of Odysseus in 3D Scenes
GAMMA decomposes long-term scene-aware human motion into generative body-marker primitives with a policy for perpetual goal-directed navigation.
Abstract
Our goal is to populate digital environments, in which digital humans have diverse body shapes, move perpetu-ally, and have plausible body-scene contact. The core challenge is to generate realistic, controllable, and infinitely long motions for diverse 3D bodies. To this end, we propose generative motion primitives via body surface markers, or GAMMA in short. In our solution, we decompose the long-term motion into a time sequence of motion primitives. We exploit body surface markers and conditional variational autoencoder to model each motion primitive, and generate long-term motion by implementing the gen-erative model recursively. To control the motion to reach a goal, we apply a policy network to explore the genera-tive model's latent space and use a tree-based search to preserve the motion quality during testing. Experiments show that our method can produce more realistic and controllable motion than state-of-the-art data-driven methods. With conventional path-finding algorithms, the generated human bodies can realistically move long distances for a long period of time in the scene. Code is released for re-search purposes at: https://yz-cnsdqz.github.io/eigenmotion/GAMMA/
How to read this
- Category
- Method: long-term scene-aware human motion generation
- Contributions
- GAMMA, which decomposes long-term motion into a sequence of generative motion primitives modeled with body surface markers and a conditional VAE
- Recursive application of the generative model to produce perpetual, arbitrarily long motion for diverse body shapes
- A policy network plus tree-based search over the latent space for goal-directed control while preserving motion quality
- Context
- Tackles character-scene interaction and navigation in the spirit of learned controllers like Starke et al. (Neural State Machine for Character-Scene Interactions, 2019), using marker-based generative primitives instead of a state machine.Builds on: Neural State Machine for Character-Scene Interactions
- Correctness
- The approach assumes recursively chained motion primitives stay stable and plausible over long horizons, with the tree search guarding quality; it reports more realistic and controllable motion than data-driven baselines, but plausibility of body-scene contact over very long durations is the thing to watch.
- Clarity
- Moderately dense; a first pass conveys the primitive-plus-policy structure, a second pass for the CVAE and the search procedure.
- How to read it
- Read once for the decomposition into recursive marker-based primitives and the goal-reaching policy; second pass on the tree-search control and how contact and quality are maintained over long sequences.
Builds on
Built upon by
Nothing yet.
Related work
- Interactive Character Control with Auto-Regressive Motion Diffusion Models 2024 / TOG
- PFPN: Continuous Control of Physically Simulated Characters using Particle Filtering Policy Network 2021 / MIG
- Composite Motion Learning with Task Control 2023 / SIGGRAPH
- PDP: Physics-Based Character Animation via Diffusion Policy 2024 / SIGGRAPH Asia
Keywords
This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →