Humanoid robots learn how hard to press: FLoRe captures contact force from robot-free demos and reaches 80% success on three contact-rich tasks
Synopsis
Everyday humanoid work such as pushing a cart, moving a box or wiping a surface needs controlled contact force, not just tracked hand motion. FLoRe makes force a shared interface across demonstration, policy and whole-body control: two handheld grippers with six-axis force/torque sensors record hand motion and contact wrenches without a robot, a diffusion policy predicts per-hand force commands from those demonstrations, and a whole-body controller tracks them by integrating the measured force error. On a Unitree G1 across three contact-rich tasks, the force-reactive policy succeeded in 80% of rollouts, against 26.7% for a position-only policy trained on the same demonstrations and run by the same controller.
Interpretation
A humanoid can now apply a commanded contact force to an object while walking and keeping its balance, rather than only reproducing hand trajectories. Earlier learned humanoid whole-body controllers either tracked motion alone or treated external force as a disturbance to resist or a stiffness to lower; direct tracking of a commanded force had been shown only on quadrupedal manipulators, using estimated rather than measured force. On a Unitree G1 across three contact-rich tasks, the force-reactive policy succeeded in 24 of 30 rollouts, or 80%; a position-only policy trained on the same demonstrations and run by the same controller with force commands disabled succeeded in 8, or 26.7%.
The approach puts force into the demonstration capture, the policy's inputs and outputs, and the low-level controller's execution. Existing robot-free whole-body demonstration interfaces recorded motion but not contact force, so imitation learning had no force supervision. Each handheld gripper carries a six-axis force/torque sensor between handle and gripper, plus a wrist camera, a pose tracker, and chest and ankle wearable trackers; the robot wears the same gripper-sensor assembly on each wrist, so sensor frame and gripper geometry match between collection and deployment.
The low-level controller outputs only joint positions and integrates the measured force error to push the position target along the commanded direction until the hand presses with the commanded force. A spring law that shifts the target in proportion to an estimated contact force cannot detect contact, and the force it produces scales with the unknown contact stiffness. In simulation the controller tracked 5 to 25 N commands per hand with a gain of 1.03 to 1.06 and about 1 N in-contact error; by comparison, SONIC plus an analytic admittance servo realized only 0.67 to 0.70 of the commanded force, with about 7.9 N in-contact error.
Perspective
The demonstrated scope is three contact-rich tasks on a Unitree G1: pushing a gantry that hangs another robot, bimanual box rotation, and board wiping, with 10 rollouts per task per policy and test conditions paired across policies; simulated force tracking is evaluated in MuJoCo on two held-out motion clip sets. A reader interested in other robot platforms, other contact objects or longer-horizon tasks should watch for replication in those settings.
On the robot the sensor still reads 1.25 to 1.50 N with nothing commanded, so deployment zeroes commands below 1.5 N and caps them at 25 N; lateral and downward commands reach lower gain and leak partly into other axes. The training contact plane is frictionless and commands are purely normal, so tangential loads never appear in training. Removing the force observation lowers success to 66.7%, which the authors describe as a smaller trend.