Reason to Behave: Achieving Human-Like Task Execution for Physics-Based Characters
Keywords: animation character control, physics-based character control, humanoid skill learning
Abstract: Two challenges have consistently been key in human-like agent research: 1) Many jobs require human-level reasoning and introspection capabilities to discern underlying patterns. 2) Controlling complex modeled humanoid characters in indeterministic black-box physical environments demands a powerful controller to exhibit human-like movement and commonsensical behavior. To the end, we introduce ''Reason to Behave", a synergistic framework combining large language models (LLMs) based introspective reasoner with an enhanced controller. The reasoner empowers the agent with extensive world knowledge and semantic insights, enhancing contextual interpretation and reasoning formulating a code-based action plan to bridge the gap between high-level instructions and the underlying simulator. The steerable controller embeds motion-phase representation into adversarial motion prior to the precise timing of diverse life-like behaviors, allowing rapid mastery over 100 semantically distinct actions, ranging from locomotion, dance, and sport to challenging specialized maneuvers, preventing mode collapse during skill learning. Without any reward-shaping or training, our character intuitively performs commonsensical behavior, excelling in many real-world tasks from navigation to more complicated challenges like room escaping and pressure plate puzzle.Videos, codes are available at https://sites.google.com/view/reasontobehave.
Primary Area: general machine learning (i.e., none of the above)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7829
Loading