A Robot Needs More Than the Right Moves to Wipe a Board

Move a whiteboard three centimeters away and a robot can keep making a convincing wiping motion without erasing anything. The hand follows its path. The eraser never presses hard enough against the board.

That was one failure in a new humanoid-robot study called FLoRe. Instead of learning only where to move, the system also learns how much force to apply. In the authors’ small whiteboard test, it erased the marked strip in eight of ten attempts, compared with three of ten for a position-only alternative.

The result points to a missing part of many demonstrations: a visible movement does not tell a robot how hard the world is pushing back, or how hard it should push in return. For a humanoid, getting that wrong can also upset its balance.

Recording the Pressure a Video Cannot Show

The researchers, affiliated with Amazon FAR and three universities, describe FLoRe in a preprint submitted on October 26. They tested a Unitree G1 humanoid on wiping a board, moving a wheeled gantry and rotating a box with both hands.

To teach these tasks, a person uses two handheld grippers equipped with cameras, motion trackers and force-and-torque sensors. A chest tracker records body movement; ankle trackers also capture the walking task. The human can collect demonstrations without operating a robot, but this is not learning from ordinary video alone. The instruments record contact as well as motion.

Those two kinds of information answer different questions. A camera can show an eraser moving across a surface. A force measurement helps distinguish an eraser pressing against it from one hovering nearby. A box’s movement shows where it went, but not the squeeze and push that made it move.

FLoRe carries both signals into execution. Its learned policy, the system that chooses actions, receives images, information about the robot’s state and measured forces at its hands. It predicts movements and desired forces. A whole-body controller then adjusts hand-position targets to bring measured force toward the requested force while coordinating the rest of the robot.

The useful instruction is no longer just “put the hand here.” It also specifies how hard that hand should press. Along directions where no force is requested, the controller can give way rather than rigidly resist an external push.

Feeling Resistance Is Not Enough

The researchers separated two possible benefits: knowing what force the hands feel, and actually sending force commands to the controller.

Across the three tasks, the complete system succeeded in 24 of 30 attempts. The position-only comparison succeeded in eight. Both used the same demonstrations, network architecture and training schedule, and ran through the same underlying controller; the comparison’s force channel was disabled.

Then the team kept the learned force prediction and force observations, but withheld the force command from the controller. Success fell to 11 of 30 attempts. The system could sense resistance, yet it could respond only through movement commands. That makes this more than a story about adding a sensor: the robot needs a way to act on what the sensor reveals.

Removing measured-force observations from the high-level policy, while retaining its force outputs, produced 20 successes in 30 attempts. The authors report statistically significant evidence for the larger drop when force commands were withheld; they describe the observation-only difference as a smaller trend. The experiment does not establish that sensing alone caused the full improvement.

This complements DexPort’s distinction between a robot hand’s visible position and its recorded commands. FLoRe addresses a further problem: even a movement command may be incomplete when the task depends on sustained contact.

Three Tasks, Not General Household Ability

The 30 attempts were ten per task, with a separately trained policy for each task and matched test conditions across the comparisons. The researchers varied loads and starting poses for the gantry, masses and angles for the box, and board position and surface for wiping.

These are useful controlled changes, not a broad test of unfamiliar homes or objects. The complete system still failed six attempts. In the gantry task, for example, a strong push coinciding with a step could cause failure. The project page still lists code as coming soon, and this author-run preprint is not an independent replication.

Nor is FLoRe the first way to use contact information. An earlier camera-free assembly system used the history of a robot’s joint movements to help guide insertion. FLoRe’s specific contribution is to connect measured forces in human demonstrations to learned force commands executed by a balancing humanoid.

The whiteboard makes the distinction unusually clear. Reproducing a gesture is not the same as reproducing its effect. A useful robot must learn not only the path through space, but the contact that makes that path accomplish something.