Ctrl-F Gives Humanoids Force-Aware Control
UC Berkeley’s Ctrl-F framework adds explicit 3D force commands to humanoid motion tracking, enabling online adaptation as object loads change. On a Unitree G1, it pushed and kicked boxes weighing up to 20 kg where a force-blind baseline failed. ArXiv.
Ctrl-F addresses a major weakness in motion imitation: copying a trajectory is not enough when success depends on exerting the right physical force.
- –Separates motion tracking from force planning through a hierarchical controller
- –Uses object-state feedback to adjust force without directly measuring mass or contact force
- –Demonstrates one interface across hand pushing, dragging, lifting, and foot kicking
- –Hardware results are compelling, but remain bounded by predefined motions, object-state estimation, and tested load ranges
- –Shows a practical path toward more robust humanoid loco-manipulation beyond force-blind policies
DISCOVERED
1h ago
2026-11-06
PUBLISHED
2h ago
2026-11-06
RELEVANCE
AUTHOR
JonathanJarvis