| Proj No. | A3218-251 |
| Title | VLA-RL – Enhancing General Robotic Manipulation with Online Reinforcement Learning |
| Summary | Existing Vision-Language-Action (VLA) models excel in robotic manipulation but struggle with generalization. VLA-RL integrates online reinforcement learning, enabling trial-and-error optimization to improve VLA adaptability to new tasks. This final year project aims to develop a language-conditioned Markov Decision Process, where the VLA policy is progressively refined based on natural language task descriptions. Additionally, we will explore using a vision-language model as a general reward function to address sparse rewards, ultimately improving the model’s ability to generalize across diverse manipulation tasks. |
| Supervisor | Ast/P Wang Ziwei (Loc:S2 > S2 B2C > S2 B2C 83, Ext: +65 67906366) |
| Co-Supervisor | - |
| RI Co-Supervisor | - |
| Lab | Internet of Things (Loc: S1-B4c-14) |
| Single/Group: | Single |
| Area: | Intelligent Systems and Control Engineering |
| ISP/RI/SMP/SCP?: |