Robot learns to write with a freely held pen using force-aware training

Researchers led by Rongqian Chen have presented Grip, Trace, Write, a simulation environment and learning framework for robotic handwriting in which the robot grips the pen instead of having it mounted, in a paper accepted to the CoRL 2026 PhysWM workshop.

A pen held this way ties together grip stability, tip motion and contact force, the authors write. Their controller has two branches, one keeping hold of the pen and one moving the arm and wrist to write, which stay linked in a closed loop. Training first learns a grasp, then copies demonstrations made with inverse kinematics, then refines the writing with reinforcement learning that rewards ink, force, time and keeping hold of the pen.

Across varied tasks, the policy completed more writing, more precisely and with better force control than behaviour cloning or inverse kinematics alone, and wrote English and Chinese text it had not seen without fine-tuning, the authors report.