This project is an improvement based on the AMP code from here https://nv-tlabs.github.io/ASE/, resulting in a system that generates physically-based human-object interaction motion.
The methodology is based on the adversarial imitation learning approach for interaction motion generation proposed in the SIGGRAPH 2023 paper, Synthesizing Physical Character-Scene Interactions. \ (Note: only the chair interaction task from the paper was reproduced, though the principles are consistent across other tasks.)
- When the input is a single-style motion clip
Multiple characters interact with the object in the same style, with limited generalizability in their motion paths.
- When the input includes multiple styles of motion clips
Multiple characters interact with the object in various styles, demonstrating good generalizability in their motion paths. Generalizability here means that when the character's initial position and orientation are
In addition to human-object interaction motion, I also explored** human-terrain interaction** (similar principles to interaction motion) with some experiments.
- Ascending stairs along a fixed path:
- Autonomous exploration on terrain:
Code for human-terrain interaction can be found here: https://github.com/budiu-39/AMP_terrain
Reproducing human-object interaction motion generation based on the paper’s theoretical foundation is relatively straightforward, though transferring this approach to a complex terrain setting has been less effective. \ Generating generalized 3D human motion for terrain-based interactions currently lacks a robust solution, making it a promising direction for further exploration!