Robot Learning based on Long-Horizon Visual-Language-Action Models

The project addresses a central challenge in embodied intelligence: enabling robots to interpret language instructions, ground them in visual observations, and execute coherent multi-step behaviors over extended temporal horizons. Recent advances in multimodal learning have led to strong progress in language-conditioned robotic control. However, long-horizon tasks remain challenging due to compounding errors, partial observability, delayed dependencies, and the need for robust subgoal reasoning, temporal abstraction, and recovery from failure. This internship focuses on learning-based approaches for building VLA systems that can support reliable sequential decision-making in complex robotic environments.