Building Common Behavior Data (CBD) — an experimental open layer connecting video, language, motion, simulation & robotics. OSS adapters & demos to develop AI.
Behavior should be portable.
What if human behavior could be represented once, then reused across robotics, simulation, and AI applications and avatar, VR?
I built Common Behavior Data (CBD) to explore that idea.
Human video
→ MediaPipe
→ Common Behavior Data
→ MuJoCo / Unity / behavior dataset
Same source behavior, multiple representations.
CBD is still experimental. I want to test where this abstraction works, where it breaks, and what a reusable behavior representation actually needs to preserve.
Open source:
https://t.co/knCw6Jk80F
Nobody holds the spray gun.
A wheeled robot rolls into an empty room and paints the walls and ceiling by itself.
The arm tracks every corner. The painter just watches.
Have succeeded to add Isaac Sim GR00T G1 adapter of CBD. It makes me perceive many insight. So I've updated CBD concept and GitHub.
https://t.co/niQQl2jRpa
With my experience and check in the past, Colab has standard security of Google Cloud and private. And it may not be reasonable for attackers to attack the environment its all data be reset after every session.
Of course we can never be too careful. file in Drive share, hard coding API keys, such actions are not recommendable, and if the organization policy is sever, other services be option.
Google Colab CLI is so exciting tools, which enables automation of ML experiment.
We can expect autonomous Colab operation by AI Agent like Claude Code/Codex...
Have updated the first CBD separated generator and adapter as ipynb which can be tested with Colab CLI. I recommend it.