1X Technologies just pulled the curtain off Redwood AI, a vision-language transformer that’s set to supercharge its NEO humanoid robots. This is a huge leap toward robots that can see, talk, and move through the world with an almost human-like finesse.
Redwood AI is a lean, mean machine, packing 160 million parameters and running entirely on NEO’s onboard GPU—no cloud needed. That efficiency is a game-changer, letting the robot tackle complex tasks in real time without the lag or privacy risks of offloading data to servers. By weaving together pre-trained language embeddings, vision tokens from a transformer, and proprioception data from the robot’s joints and forces, Redwood builds a deep, nuanced understanding of its surroundings. “Redwood is tailored for the humanoid form factor and capable of performing end-to-end mobile manipulation tasks like retrieving objects for users, navigating around, and opening doors,” 1X shared in their announcement.
- With this set, based on the legendary Transformers Autobot, adults build LEGO Optimus Prime figure that converts from robot to truck and back
- Relive the Transformers saga with accessories, including the ion blaster, Autobot Matrix of Leadership, an Energon axe, Energon cube and jetpack
- This unique collectible robot model makes a great home décor piece - top off the build with a unique display plaque with Optimus Prime facts
To get Redwood up to speed, 1X fed it a rich dataset gathered from NEO and its predecessor, EVE, across their offices and employees’ homes. This real-world training lets Redwood handle the chaos of everyday life, from grabbing unfamiliar objects to weaving through new spaces. Unlike old-school robotic systems that treat walking and grabbing as separate jobs, Redwood fuses them into one fluid system, letting NEO lean into a door or crouch for a toy with a natural, almost human grace. “Redwood is among the first VLAs to control locomotion jointly with manipulation, enabling bracing and leaning behaviors,” 1X noted.
By incorporating failure rollouts into its training, the model refines its cognitive predictions, avoiding the trap of overfitting to perfect scenarios. Successful demonstrations guide its action diffusion heads, while failures sharpen its understanding of the world. This dual approach ensures NEO can adapt to new challenges, like grasping a bottle it’s never seen before. “It learns from both success and failure to get smarter with every task!,” said one social media post.
NEO’s latest reinforcement learning controller, powered by Redwood, enables natural walking, kneeling, and even stair climbing using stereo vision. This unified controller, a milestone in humanoid robotics, allows NEO to navigate tight spaces—like squeezing between a sofa and a coffee table—where wheeled robots would falter. “Having complete and unfettered access to all parts of the world is important for NEO to accomplish tasks in home environments,” 1X emphasized.
Features like emotive ear rings and a three-speaker audio system enhance communication, making interactions feel natural. “NEO Gamma’s design opens the door to start internal home testing—a first step in creating fully autonomous humanoids,” 1X stated.
