FlowDAgger: Human-in-the-Loop Adaptation of Generative Robot Policies in Latent Space

Technical Analysis is a sample- and compute-efficient method for adapting frozen generative robot policies from human interventions in latent space. Each human expert action is mapped to the noise that would have produced it under the frozen base policy, using reverse-time integration followed by local refinement. High-level Analogy: Imagine a brilliant chef who has a massive recipe book full of amazing dishes (the robot's pre-trained knowledge or 'base policy'). This chef is generally very good, but sometimes a customer (a human operator) has a specific, tricky request or a unique ingredient that isn't quite right for the standard recipe. Instead of asking the chef to rewrite entire recipe chapters (which would be slow and might mess up other recipes), the customer gives a small, precise feedback: 'Make it a bit sweeter and less spicy' (the human's corrective action). FlowDAgger is like a super-smart assistant who observes this feedback. This assistant doesn't change the chef's main recipe book. Instead, they figure out exactly what tiny, internal adjustment to the core ingredients (the 'noise vector' in the latent space) would have led to that perfect taste the customer wanted. The assistant then learns this specific 'adjustment formula' for similar situations. So, the next time a similar picky customer appears, the assistant automatically tells the chef, 'Add a tiny pinch of sugar and reduce the chili just a bit' (applies the learned noise vector), steering the chef's…