1X Redwood AI

2025-06-10 · Publisher : 1X · Domain : Robotics

1X unveils Redwood, a 160-million-parameter vision-language model that drives end-to-end mobile manipulation for its NEO Gamma humanoid at home, entirely on the onboard GPU.

What changes

160 million parameters running on NEO's onboard GPU.
One of the first VLAs coordinating locomotion and manipulation.
One architecture for EVE and NEO robots, with a diffusion policy decoder.
Trained on data collected in 1X offices and employees' homes.

### Background On June 10, 2025, 1X introduced Redwood AI, the model it plans to deploy in homes with its NEO Gamma humanoid. Redwood is a vision-language transformer tailored to the humanoid form, performing end-to-end mobile manipulation: fetching objects, opening doors, navigating the home. ### Specifications - **Size**: 160 million parameters, a deliberately compact model. - **Execution**: fully on NEO's onboard GPU. - **Whole body**: per 1X, one of the first VLAs to control locomotion jointly with manipulation, enabling bracing and leaning during a task. - **Architecture**: fuses pretrained language embeddings, vision tokens and proprioception (joint positions and forces), decoded into EVE or NEO actions by a diffusion policy. - **Data**: teleoperated and autonomous episodes from EVE and NEO, collected in 1X offices and employees' homes. ### Why it matters Redwood shows an alternative to large VLAs: a small, specialized, onboard model trained on real home data. For a consumer robot, local execution matters for both latency and privacy. ### Key takeaway 1X prepared NEO's commercial launch with an in-house model designed for the home; preorders opened in October 2025.

1X Redwood AI

What changes

Links