A robot policy can understand the scene and still fail because its training data is thin, its controller is late or its motors cannot keep up. Physical Intelligence's PI0.5 and NVIDIA's GR00T N1 establish how vision, language, robot state and fast action fit together. Waymo's EMMA planner tests end-to-end control where a plausible mistake is dangerous, before K-Scale moves the same concerns into an open humanoid stack. CloudChef brings the system into a hot, variable workplace where thermal sensing and human recovery matter as much as the model. A control-systems case traces bad motion through stale sensor data and logging pauses. Voice comes only after that machinery is visible. Scout's stage fall makes the limit plain: a clean language interface can choose a policy, but it cannot repeal physics.
Why first: Vuong and Springenberg define the problem behind the rest of the pack. Language and vision are not enough when a model must produce fast physical actions. They also put the less glamorous constraint first: before architecture or deployment, someone has to build the robot-data operation.
Why second: Vuong and Springenberg explain why a language model needs an action layer. Brundyn and Jhunjhunwala open that layer up: a slow planner, a fast controller and a decoder for each robot body. Their data pyramid also answers the scarcity problem the first talk introduces.
Why here: Brundyn and Jhunjhunwala describe a general architecture. Hwang puts the same idea into autonomous driving, where a plausible action is not enough. EMMA makes intermediate decisions visible, then uses generated weather and lighting changes to test the planner under conditions that are hard to collect safely.
Why here: Hwang shows why simulation matters when a bad decision is dangerous. Mo makes simulation the developer's entry point. KS Sim and the physical robot share an interface, so moving a controller onto the hardware means changing its target rather than rewriting it.
Why here: Mo gives developers a general robot stack. Abraham shows what changes when one enters a particular workplace. Cooking needs thermal sensing, recipe state and a person ready for the odd case. General hardware becomes useful through evidence about this job, not through generality alone.
Why here: Abraham separates good perception from slower physical execution. Garg explains how the same gap can make a sound policy look broken. Stale sensor data, queued commands, logging pauses and kernel scheduling all turn correct computation into bad motion.
Why here: Garg shows that a few milliseconds can change physical behavior. Marafioti applies that lesson to the person facing the robot. Reachy Mini combines repairable hardware, partial transcription and faster speech generation so the interaction begins while a remote model would still be waiting.
Why last: Marafioti makes conversation with a robot faster. Subramani shows what conversation can and cannot control. Her agent chooses an existing policy, while the policy decides the movement. Scout falling over on stage is the useful ending: a clean language interface does not remove the physical system underneath it.