- Published on
1X Reveals the RL Core That Gives Its NEO Humanoid Fluid, Whole-Body Mobility
- Authors
- Name
- Humanoids daily
- @humanoidsdaily
Just one day after announcing Redwood, its unified AI model for the NEO humanoid robot, robotics firm 1X has pulled back the curtain on the technology that allows its robots to move. In a new technical blog post and a series of demonstrations, the company detailed the reinforcement learning (RL) system that underpins NEO's physical agility.
This isn't just a walking algorithm; it's a foundational mobility core that blends a wide range of behaviors—walking, running, sitting, squatting, and getting up from the floor—into a single, controllable intelligence.
"Today, we're unveiling the RL core that makes Redwood possible," announced 1X CEO Bernt Børnich. He described it as a system that "brings us closer to the end game of autonomous agents that explore, adapt, and learn directly from the real world."
Redwood AI | Mobility Reinforcement Learning
Unifying the Body with a Single Controller
The central challenge 1X aims to solve is the fragmentation common in robotics, where locomotion (legs) and manipulation (arms) are often treated as separate domains. This can limit a humanoid's ability to perform complex, whole-body tasks. Eric Jang of the AI team at 1X noted that many roboticists still use humanoids as "bimanual arms on a pedestal."
1X's goal is to move beyond that paradigm. The new RL controller is designed to give the high-level Redwood AI access to the robot's full physical capabilities, enabling more dynamic and fluid movements necessary for navigating unstructured human environments.
To achieve this, 1X employs a two-stage process. First, it uses motion-capture data from humans to learn what natural movement looks like. However, simply replaying these motions isn't useful for a general-purpose robot that needs to react to commands. The key innovation is a system that can translate a high-level command—either from the Redwood AI or a human teleoperator—into a fluid, stable, and steerable action.
A high-level "kinematic planner" generates a target trajectory that resembles human motion, and a low-level RL controller then figures out the physics of how to track that trajectory while keeping the robot balanced. A demonstration of the Redwood AI using this system to autonomously dribble a soccer ball highlights the steerable, reactive nature of the controller.
Climbing Stairs with Vision Alone
One of the most significant capabilities showcased is NEO's ability to ascend and descend stairs. For a robot intended for home environments, mastering stairs is a critical milestone for providing "unfettered access to all parts of the world," as the company puts it.
Notably, NEO accomplishes this without the aid of lidar or dedicated depth sensors, which are commonly used in legged robotics for mapping terrain. Instead, the robot relies purely on a pair of stereo RGB cameras. The RL controller infers depth and step height directly from the visual feed, fusing it with the robot's own sense of motion (proprioception) to plan its steps.
According to 1X, training the system with extensive domain randomization in simulation has made the controller robust enough to handle stairs of varying heights and even to side-step while on them.
Safe by Design
The demonstrations feature the NEO robot moving with a gentle, compliant quality—a physical characteristic 1X sees as crucial to its strategy. This trust in the hardware was put on full display in a video (see above!) showing the robot walking upstairs while Dar, the company's head of design, sits calmly reading just inches away.
He later highlighted the importance of this design choice in a comment about the scene. "Worth noting the importance of a safe robot here," he wrote. "If NEO wasn’t extremely lightweight, compliant and soft— we’d never be able to shoot this scene."
This combination of a soft, lightweight body and a sophisticated, unified mobility controller represents 1X's bet on how to build a truly general-purpose humanoid: create a physically safe platform, then empower it with an AI that can learn to use its entire body to navigate and interact with the world as a whole.