In this thesis, we propose a vision-based deep learning control framework that directly generates joint trajectories of a robot manipulator from RGB camera images and joint states. The goal of the framework is to map high-dimensional visual observatio...
In this thesis, we propose a vision-based deep learning control framework that directly generates joint trajectories of a robot manipulator from RGB camera images and joint states. The goal of the framework is to map high-dimensional visual observations to physically plausible robot motions in a stable manner while preserving as much as possible the exploration capability and reliability of traditional motion planners. To this end, several vision-based control modules with different levels of abstraction are designed, and a hierarchical controller that combines these modules is constructed and systematically compared in both real and simulated robotic environments.
First, we design a ResNet-based key-point predictor that estimates task-relevant object key-points from camera images, and we construct a module that plans collision-free joint trajectories using a tree-based motion planner that treats the predicted key-points as target states. Second, we implement a ResNet-based joint state predictor that takes RGB images and the current joint state as input and directly regresses the joint state one step ahead, thereby learning an end-to-end visuomotor mapping. Third, we introduce a Transformer-based joint trajectory prediction model that uses visual observations and the current joint state to predict a multi-step joint state sequence in a single forward pass, enabling sequence-level control policies that capture long-term dependencies. Finally, we construct a hierarchical vision-based controller in which a high-level ResNet-based key-point module proposes task-level goal states and a low-level learned trajectory generator samples detailed joint trajectories, and we analyze the performance gap between this hierarchical controller and single-stage models.
The proposed methods are evaluated on three manipulation tasks: a block–cup color-matching task on a real Kinova Gen3 Lite robot and a mug upright and box insertion task in a PyBullet–Panda simulation environment. In the PyBullet–Panda setup, an RRT-based motion planner is used to automatically generate collision-free joint trajectories for diverse initial object placements and goal configurations, and these trajectories serve as training data. In the Kinova Gen3 Lite environment, a human operator performs the tasks via teleoperation, while camera images, joint values, and gripper states are synchronized and recorded. Each model is then trained in a supervised manner to imitate the recorded joint trajectories.
Experimental results indicate that the sequence prediction–based Transformer model outperforms the single-step joint state predictor in terms of long-horizon tracking accuracy and final goal-reaching success rate. In particular, for complex manipulation tasks such as the mug upright and box insertion scenario, where multiple contact events occur and trajectories are long, the hierarchical controller achieves higher task success rates and produces more stable trajectories than a non-hierarchical Transformer-based controller.
Overall, this thesis (i) compares and analyzes key-point–based high-level motion planning and joint sequence–based low-level control within a unified framework, (ii) constructs datasets in simulation and uses them to quantitatively evaluate the performance of various deep learning models, and (iii) transfers models evaluated in simulation to a real robot arm and demonstrates that they can be executed successfully without architectural changes. These results contribute to the design of robust vision-based manipulator control systems that can be extended to a broader range of objects and tasks in future work.