UC Berkeley

This dissertation presents a series of statistical learning methods to improve the overall

performance for industrial robots, which are always required to perform accurate and fast

motions in sensing limited scenarios, and/or changing environments. Statistical methods

emphasize on system uncertainties and adaptation to the environment by learning from

the experience. Four aspects have been studied in detail including robust visual servo,

optimal trajectory planning, intelligent system identication for feedforward compensation,

and optimal gain tuning.

For visual servo requiring high precision and high speed but suffering from limited sensing

capability and system uncertainties, a statistical learning approach based on sensor fusion

and robust vision is adopted for sensor data filtering and on-line parameter adaptation.

This method can compensate for the large sensing latency and estimate system parameters

simultaneously. The algorithm is tested on an 6-axis industrial robot performing precision

glass handling task but with only limited quality vision sensors. Trajectory planning is

another fundamental problem for industrial robot, especially for robots operating in extremely

limited workspace. In order to generate obstacle-free trajectories with shortest cycle

time, a statistical learning method, Probabilistic Roadmap (PRM), is first used for optimal

trajectory planning. An optimization problem is then formulated based on spline parameterization.

The planning algorithm is tested on a 3-axis industrial robot, and the result shows

the efficiency of the planning method. Statistical learning approaches are also studied for

the precision motion control and vibration suppression for industrial robots. Model-based

controller for parallel robots is always hard to design due to the complexity of the robots. A

model based feedforward controller is designed with dynamic identification with a regression

performed using data generated by experiments with designed trajectory. Also, an experiment

based guided gain search is proposed with direct end-effector sensing for vibration

suppression. An optimization scheme is utilized to guide the searching direction with experiments

and data learning. It is proved by experiments that the tracking error and vibration

are greatly reduced on an 8-link wafer handling robot.