On 21 October, we shared an update on RSO Pose, our movement prediction model. The work focuses on terrain-aware locomotion with a small footprint, with potential uses for animators, independent game developers and roboticists.
Our July introduction described how RSO Pose learns from human motion capture to generate walking, running and turning. This update brings the ground into focus: how a continuing sequence of predicted poses fits the scene a character moves through.
From a pose to a step
The controller uses recent movement and a requested direction to predict the next pose. That pose then becomes part of the history for the following prediction. A step develops through this repeated process, alongside the turns and changes of pace requested by the person controlling the character.
Ground contact gives us something concrete to examine. Does the foot meet the surface? Does it stay in place while it carries the step? Does the body continue naturally into the next movement? We use control and contact adjustments alongside the learned poses to work on those details in simulation.
Reading the ground ahead
Our terrain-sensing studies explore how information about the surface ahead could guide movement prediction. The blue scanning treatment shown in the site’s studies illustrates that research direction.
A compact system for interactive work
A small footprint matters for the applications we have in mind. An independent developer needs room for movement alongside the rest of a game; an animator needs a controller that can respond while a scene is being explored. These are practical reasons to pursue compact motion systems.
The controller runs in simulation. Taking it onto a physical robot will require testing balance and contact forces on the hardware.
From the archive. The date refers to the period described.
Watch the October 21 RSO Pose demonstrations