FLEX-π: A MULTI-STREAM WORLD-ACTION MODEL WITH COMPUTE FLEXIBILITY

Technical Analysis is a WAM that predicts 3D geometry and object-centric DINO semantics alongside RGB. Per-stream dropout with cross-modality forcing lets a single trained checkpoint run on any subset of these streams, from a fast action-only mode to full joint generation. High-level Analogy: Imagine you're a super-smart chef trying to prepare a complex dish. Older robot chefs might only be able to see the color of ingredients. They might struggle to tell the difference between a flat red pepper and a round red tomato just by color. FLEX-π is like a chef who not only sees the color (RGB image), but also feels the exact shape and texture (3D pointmap), and instantly knows what each ingredient is (DINO semantics – 'this is a bell pepper, this is a cherry tomato'). All this information is processed together in one swift thought process in their brain. Here's the clever part: Even if you temporarily 'blindfold' this chef (e.g., remove the 3D depth sensor), they've learned so well during training that they can still accurately imagine the shape of the ingredient from its color and type. And if they just need to quickly grab something (action-only mode), they can do it faster than other chefs, but the training they received by 'imagining' everything thoroughly makes them much more skillful overall. This makes them incredibly adaptable and efficient! Motivation of the Work The core motivation behind FLEX-π is to enhance the already strong capabilities of World-Action Models (WAMs)…