Robustness and
Few-Shot Adaptation
A research project on making Vision-Language-Action models robust to distribution shifts and
able to adapt to new environments and tasks from only a few demonstrations
Overview
This project studies the robustness and few-shot adaptation of Vision-Language-Action (VLA) models, which map visual observations and language instructions directly to robot actions for embodied manipulation.
Although VLA models show strong generalist ability, their performance often degrades under distribution shifts—changes in environment, scene layout, or task that differ from the training data. Retraining or fully fine-tuning these large models for each new setting is costly and impractical, creating a need for efficient adaptation with only a handful of examples.
To address this, we reproduced representative VLA models and built an end-to-end inference and evaluation pipeline on the LIBERO and RoboTwin manipulation benchmarks. Using the empirical failure patterns we observed, we explore training-free, memory-based, context-conditioned, and residual adaptation mechanisms that let a policy adjust at test time without committing to a fixed architecture.