A mobile robot placed bread on a plate in 19 of 20 trials after 25 minutes of on-robot fine-tuning, compared with 10 of 20 trials for its original controller, researchers report. The result came from one task in a new study of how robots can adapt their driving and arm movements with limited human guidance.
The work, called HiRe-MoMa, was submitted to arXiv as a version 1 preprint on Oct. 14. The researchers tested it on a TIAGo++ mobile manipulator, a robot that combines a wheeled base with two arms.
HiRe-MoMa combines learned movement corrections with limited human guidance
HiRe-MoMa leaves the robot’s pre-trained controller fixed and learns small corrections that are added to its actions. Instead of producing one combined correction for the whole robot, it uses one module for the base and another for the arm. The arm module receives the base correction first, because where the robot parks changes what the arm can reach.
Human operators’ corrections while remotely controlling the robot made up less than 20% of the training data, the authors say. During learning, the robot also used an action-scoring model that estimates which moves are more promising. The researchers trained the scoring model to give higher scores to human-corrected moves than to the robot’s own alternatives, so it could use occasional guidance while continuing to explore rather than copying every human movement.
Place Only and sweeping reached 90% or more; Pick Place used partial credit
In the Place Only task, the robot began with bread in its gripper, drove to a table and put the bread on a plate. Across 20 trials, the frozen Diffusion Policy controller succeeded 50% of the time; the full HiRe-MoMa system reached 95% after 25 minutes of real-world fine-tuning. On a separate sweeping task, success rose from 45% to 90% after 30 minutes.
The longer Pick Place task remained harder: its reported score rose from 22.5% to 40% after 40 minutes. That score gives partial credit for grasping the bread and full credit for placing it, so 40% does not mean the robot completed 40% of trials from start to finish.
One robot ran 20 trials per task
The researchers ran 20 trials per task on one TIAGo++ platform, with the robot’s torso and head fixed and starting positions and objects randomized. A result on one platform does not establish how reliably the method will work across other robots, homes or workplaces.
Sources
Photo: PAL Robotics, via Wikimedia Commons, CC BY-SA 4.0. The image shows the TIAGo++ model; it is illustrative and is not from the study.