Stanford's HomeBody runs a humanoid on a raw VLM, no VLA layer

No VLA training, no per-environment data — the VLM calls motor skills directly. Plus: Meta patches a Muse flaw that could expose your VM's files.

Nowline SEP 28 12:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • A raw VLM drives the humanoid — no learned action policy

    Stanford and Caltech's HomeBody wires GPT Astra straight to five composable skills — navigate, pick, place, open-drawer, and pick-from-drawer — on a Unitree G1, skipping the usual System-2 → VLA → controller stack. It explores an unseen room first to build a Real2Sim spatial memory (SLAM + Isaac Sim), then reasons over that to run multi-room, multi-object cleanups with no environment-specific training, locally on an RTX 4090 laptop. The transferable idea for any agent builder: a solid skill library plus persistent memory can stand in for a trained policy layer. Project page and GitHub are live.

  • Meta patches a Muse flaw that could reach your cloud VM

    A researcher found a vulnerability in Meta's Muse — the agent that hands each user a dedicated cloud VM — that could let an attacker reach that VM's contents, including emails and files. Meta says it's fixed and has hardened Muse's safety warnings. If you've connected real accounts to a Muse VM, treat anything from before the fix as potentially exposed and rotate those credentials.