VioLA: Learning Generalist Humanoid Control Policies from Human Data
VioLA trains a generalist humanoid policy mostly on human motion and follows real-robot instructions without task fine-tuning.
VioLA is a generalist humanoid policy that predicts body and hand motion latents executed by pretrained controllers, whose encoders place human and robot motion in the same latent spaces. Training uses 140.6 million frames, 93.2% of them human. On a real robot it reaches 100% zero-shot locomotion success, versus 16.7% for GR00T N1.7 and 0% for Psi0, and 88.6% manipulation success without task-specific fine-tuning. The approach works across two VLA backbones and one world-action model; code and checkpoints will be released.
- Policy predicts body and hand motion latents, not joint commands
- Training uses 140.6 million frames, 93.2% from humans
- 100% zero-shot locomotion versus 16.7% for GR00T N1.7 and 0% for Psi0
- 88.6% manipulation success without task-specific fine-tuning
- Same approach works across two VLA and one world-action backbone
- AI models
- VioLAGR00T N1.7Psi0
From the original article
1 min read at arxiv.orgTeaching a humanoid to follow instructions with its whole body runs into two obstacles. Its action space is large and tightly coupled: legs, arms, and fingers must move together while the robot keeps its balance, which makes joint-level actions hard to learn.
By Mert Albaba · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.