Teaching robots to perform complex tasks based on our instructions is a huge challenge. Imagine trying to explain to a robot how to make a cup of coffee – it’s not just about the words you use, but also the robot's understanding of its environment. Traditional methods require a massive amount of annotated data, like detailed transcripts of every step, which is time-consuming and expensive. New research introduces FLARE, an innovative approach that allows robots to learn from significantly fewer examples by incorporating what they see. FLARE uses a “Multi-Modal Planner” that combines language instructions with the robot's visual perception to create an initial plan. So, instead of blindly following a recipe, the robot uses the context of its surroundings to generate a more sensible sequence of actions. But what happens when the robot encounters unexpected variations in language or the environment? This is where FLARE's “Environment Adaptive Replanning” comes in. Let's say the robot is told to 'put the mug on the coffee maker,' but it's been trained on 'coffee machine.' Instead of getting stuck, FLARE uses visual cues and semantic similarity to figure out that the 'coffee machine' is probably the same thing as the 'coffee maker.' This allows it to adapt its plan on-the-fly, making it much more robust to the real world's ambiguities. Experiments on the ALFRED benchmark, a standard test for instruction-following robots, show that FLARE significantly outperforms existing methods, even when trained on just a fraction of the data. While FLARE still relies on some training data, it represents a huge leap toward robots that can learn efficiently and adapt to new situations, paving the way for truly helpful robotic assistants in our homes and workplaces.