Rather than being subject to odometry drift, the robot plans deliberate contacts with walls and obstacles in order to re-localise itself, and reaches its destination more often than the benchmark methods.
In a blind environment, a humanoid robot quickly loses track: its odometry drifts and uncertainty about its position accumulates. TAPNAV tackles the problem through active tactile perception: the robot will deliberately touch the structures around it to readjust its position estimation.
Plan where to touch
The system follows a probabilistic estimation of the pose from odometry, inertial measurement unit and tactile contacts. A global planner chooses routes that keep uncertainty within acceptable limits by providing opportunities for correction by touch. A local planner then selects the palpation gestures that provide the most information. A whole-body controller coordinates walking and hand contact.
Tested in simulation and on a Unitree G1, with various floor plans and obstacle shapes, TAPNAV achieves a lower estimation error and a better success rate than reference methods. For the authors, physically interacting with the environment may be enough to provide reliable landmarks where vision is lacking.
Read the full article on the original source
Go to the source →