Skip to main content
Back

UC Berkeley's Ctrl-F Teaches Humanoid Robots to Push 20 kg Loads

Berkeley researchers give humanoid robots a force command channel, letting a Unitree G1 drag, push and kick 20 kg loads that a motion-only tracker cannot budge.

  • Berkeley's Ctrl-F adds a 3-D force command on top of humanoid motion tracking at each hand or foot.
  • A low-level tracker is trained object-free by applying random wrenches and matching force commands to the robot.
  • A high-level planner picks the force online from object state plus proprioception, no force sensor needed.
  • On a Unitree G1 it drags, pushes, raises and kicks loads up to 20 kg where the SONIC baseline fails.
  • Planner adapts mid-task when weight is added, keeping dragging pace and wrist height stable.
  • Code coming soon; paper and browser demos available on the project site.

Humanoid robots can reproduce human motion while still failing at basic physical work. A pose target can place a robot’s hand against a box, but it cannot specify how hard the hand should press. UC Berkeley researchers address that gap with Ctrl-F, a control framework that combines whole-body motion tracking with explicit 3-D force commands.

In the team’s paper, a Unitree G1 using Ctrl-F drags carts, pushes boxes, raises bins, and kicks loads weighing up to 20 kg. SONIC, the force-blind motion-tracking baseline used for comparison, often maintains the requested pose without moving the object.

Pose tracking loses its grip

Whole-body motion controllers learn reference trajectories that specify where the robot’s joints should move. Physical interaction adds another requirement: the robot must apply enough force through its hands or feet to overcome gravity, friction, and the object’s inertia. Without that information, hands can slide off a box, a loaded cart can slow the robot, and contact can destabilize a standing kick.

Earlier systems have supplemented motion tracking with object targets such as contact points or desired geometries. Ctrl-F exposes the interaction force directly. Its policy receives a reference motion plus a 3-D force vector for each designated contact body, such as a hand or foot.

Ctrl-F training pipeline with a low-level motion-force tracker and a high-level force planner
Ctrl-F trains an object-free motion-force tracker first, then freezes it while training an object-aware force planner.