The rise of robots is becoming unstoppable judging by how much effort and money has been invested in this Robotics field so far just these years. Carl Frey and Michael Osbourne in Oxford University released a paper in 2013 and claimed that around 47 percent of current jobs would be automated in the next two decades. But planning robot motion still remains a major problem in Robotics regardless of countless approaches proposed in multiple aspects trying to solve it. TrajOpt(Trajectory Optimizer) is a state-of-art optimization-based software framework for planning robot motions. TrajOpt generates trajectory through constrained sequential convex optimization given several initial guesses, meaning TrajOpt would focus on finding the local minimum guided by an initial guess. However, depending on the complex environment and robot mechanical structure, it sometimes would suffer from being stuck in the local minimum which is not a feasible trajectory. However, BiRRT(Bidirectional Rapidly exploring random tree) is probabilistic complete. BiRRT is a sampling-based method. It has been widely used due to its property, probabilistic completeness. But without using any smoothing techniques, the trajectory generated by BiRRT mostly is inexecutable on the real robot. The objective of proposing this work is to use the sample-based method to enable the TrajOpt become probabilistic complete, which guarantees that considering the solution being present the planner has the capability of acquiring the optimized trajectory. I also intend to experimentally evaluate the performance of this improved method in the simulation called Gazebo and on the real Atlas robot over a wide range of environmental settings.