Robot AI model
Hume
Hume combines value-guided VLA action selection with a faster reactive denoising policy, demonstrated on WidowX, Franka Panda and AgiBot G-1. The repository uses MIT for its own material and retains separate terms for inherited code and models; a complete checkpoint license is not established.
Records last verified Oct 13, 2026 · Send a correction
- Organization
- Shanghai AI Laboratory / Hume collaboration
- Type
- VLA
- Released / introduced
- May 27, 2025
- Public weights
- Public weights
- Licence / scope
- Not established
- Parameters
- Not disclosed
- Base model
- Not disclosed
- Humanoid association
- Humanoid demonstrated
- Inputs
- vision, language, proprioception
- Outputs
- continuous action chunks
Release dates can mark an announcement or paper introduction. Public checkpoints, code licences and weight terms are distinct; follow the publisher’s links for the stated scope.
Documented embodiments
Source-associated hardware; this list does not establish plug-and-play compatibility or commercial robot capability.
- WidowX 250s
- Franka Emika Panda
- AgiBot G-1
Project spelling; the exact catalog generation is not unambiguously identified.
Disclosed training data
- LIBERO →
- BridgeData (bridge_orig LeRobot conversion; version not established)
- Fractal
More robot AI models
Evidence
Sources
- [s1] Hume: Introducing System-2 Thinking in Visual-Language-Action Model ↗
Hume research team · paper · published May 27, 2025 · accessed Oct 13, 2026
- [s2] Hume project: method and robot experiments ↗
Hume research team · official · accessed Oct 13, 2026
- [s3] Hume official repository and checkpoint instructions ↗
Hume research team · official · accessed Oct 13, 2026
- [s4] Hume paper: affiliations, model architecture, training and robot experiments ↗
Hume collaboration · paper · published Jul 8, 2025 · accessed Oct 13, 2026