Humanoid Robotics Technology
1X previously developed an autonomous model capable of merging multiple tasks into a single goal-conditioned neural network. However, in small multi-task models (<100M parameters), adding data to correct one task’s behavior often negatively impacts other tasks. While increasing the model’s parameter count can reduce this issue, it also lengthens training time, slowing down the process of identifying which demonstrations improve robot behavior.
To iterate quickly on data while building a versatile robot capable of many tasks with a single neural network, 1X aimed to separate the quick improvement of task performance from the integration of multiple capabilities into one network. They achieved this by creating a voice-controlled natural language interface that chains short-horizon capabilities across several small models into longer sequences. With human guidance in skill chaining, this approach enables the long-horizon behaviors demonstrated in the video.
While humans can easily handle long-horizon chores, chaining multiple autonomous robot skills in sequence is challenging because each successive skill must generalize to the slightly varied starting positions resulting from the previous skill. This complexity increases with each additional skill; the third skill must accommodate the variations in outcomes from the second skill, and so on.
For users, the robot appears capable of performing many tasks via natural language commands, with the underlying complexity of multiple models abstracted away. This approach allows 1X to gradually merge single-task models into goal-conditioned models. Single-task models also provide a reliable baseline for shadow mode evaluations, where new model predictions are compared against an existing baseline during testing. Once the goal-conditioned model aligns well with single-task model predictions, 1X can transition to a more powerful, unified model without altering the user experience.
Using a high-level language interface to direct robots offers a novel data collection experience. Instead of using VR to control a single robot, an operator can use high-level language to direct multiple robots, with low-level policies executing the specific actions to achieve those high-level goals. Since high-level actions are infrequent, operators can control robots remotely, as demonstrated in the video.
SOURCE: 1X
