Video: Sweater-wearing humanoid robot gets brain upgrade to clean, cook autonomously

In February, the San Francisco Bay-based 1X Technologies launched NEO Gamma, a next-gen humanoid robot for household chores.

Intelligent humanoid control

Redwood is a compact yet powerful 160 million-parameter vision-language model designed for humanoid autonomy in home environments.

Running entirely onboard NEO Gamma’s embedded GPU at 5Hz, Redwood enables seamless end-to-end mobile manipulation, combining perception, locomotion, and control. It integrates pre-trained language and vision models with proprioception data to generate robot actions via a diffusion policy, allowing NEO to perform tasks like grasping unfamiliar objects, opening doors, and retrieving items while navigating dynamically around obstacles.

Unlike traditional systems that separate locomotion and manipulation, Redwood handles full-body coordination, simultaneously predicting arm, leg, pelvis, and walking commands. This multi-contact capability lets NEO brace against surfaces and execute complex household tasks with improved payload capacity.

According to 1X Technologies, Redwood also enables mobile bi-manual manipulation, training navigation, and manipulation jointly to optimize positioning for object handling. Its ability to generalize stems from diverse real-world training across offices and homes using both successful and failed demonstrations.

The model is further enhanced through “cognitive” prediction targets, such as tracking hand and object locations in image space, which improve grounding and adaptability.

Voice control is supported via offboard natural language processing that translates user commands into Redwood’s action space.

1X claims Redwood’s design emphasizes robustness, efficiency, and autonomy, making NEO suitable even in offline or low-connectivity environments like basements or campsites.

Dexterous robot motion

1X Technologies has developed a comprehensive Reinforcement Learning (RL) controller to expand the mobility and dexterity of its humanoid robots.

The controller enables full-body autonomy across complex tasks, supporting walking in any direction, climbing stairs, sitting, kneeling, and getting up from the floor—all crucial for navigating real homes.

Unlike conventional systems that isolate manipulation and movement, this unified controller bridges locomotion with whole-body control, powered by stereo RGB vision for accurate floor-height estimation and terrain adaptation.

The system uses a two-stage architecture: a high-level kinematic planner converts commands into human-like motion trajectories, while a low-level RL controller translates these into precise, contact-aware joint actions.

The architecture ensures fluidity, robustness, and steerability, even during task transitions. It supports side-stepping in tight spaces and climbing irregular stairs using domain-randomized training for real-world robustness.

For tasks near the ground, the controller enables NEO to safely sit, kneel, or lie down, then return to standing, expanding its reach and versatility.

Integrated with Redwood AI, the controller allows vision-driven prediction of joint and pelvis targets for tasks like soccer ball dribbling.

Together, Redwood and the RL controller form a full-stack system that enables NEO to explore and interact with the entire physical state space of a home environment.