This repository curates a comprehensive collection of research papers, articles, tutorials, and projects focused on Behavior Foundation Models (BFMs). BFMs aim to learn generalized behavioral priors from extensive, diverse datasets, enabling adaptation to a wide array of downstream tasks, particularly in robotics and character control. It serves researchers and engineers seeking to understand or leverage these advanced models.

How It Works

BFMs are developed through distinct pre-training methodologies. Forward-backward representation learning utilizes reward-free transitions to construct forward and backward embedding networks, which are then combined with reward functions for policy inference. Goal-conditioned learning relies on extrinsic rewards and large-scale human data, while intrinsic reward-driven learning employs self-supervised tasks to generate internal rewards. Adaptation strategies include common fine-tuning techniques like full fine-tuning (FFT) and low-rank adaptation (LoRA), as well as latent space adjustments. Advanced adaptation also involves hierarchical control, where high-level planners (e.g., LLMs, diffusion models) decompose abstract goals into subtasks for the BFM low-level controller.

Highlighted Details

• Pre-training Approaches: Forward-backward Representation Learning, Goal-conditioned Learning, Intrinsic Reward-driven Learning.
• Adaptation Strategies: Full Fine-Tuning (FFT), Low-Rank Adaptation (LoRA), Latent Space Adaptation, Hierarchical Control via LLMs/Diffusion Models.
• Notable Datasets: Humanoid-X, PHUMA, Motion-X++, Motion-X, HumanML3D, AMASS.
• Key Research Areas: Humanoid control, whole-body control, motion tracking, character animation, imitation learning.

Maintenance & Community

This repository is a curated list and does not appear to have active maintenance or community channels beyond the cited survey paper.

Licensing & Compatibility

No licensing information is provided for the repository itself. Individual papers linked within the list will have their own respective licenses and publication terms.

Limitations & Caveats

As a curated list, this repository does not present a deployable system. The adoption of BFMs themselves may be subject to the computational requirements, data availability, and specific limitations detailed within the individual research papers.