ROBNEWS
A new research framework called APPROVE turns a large language model into a robot programming partner, then saves the blocks the user trusts for later.
Picture the recycling line at a small community center. A volunteer points an arm at a bin of mixed bottles and cans, describes in plain language what should go where, and waits. The arm does not move yet. On the volunteer's laptop, a chain of colored blocks spells out the AI's plan: pick up, check label, route to plastic, route to glass. The volunteer deletes one step, drags a new one in, and clicks approve. The arm runs.
That is the experience a new research framework, APPROVE, is built around. The acronym expands to AI-Powered Programming for Robots with Visual End-User Feedback, and the authors describe it as a way to put a large language model, the kind of AI behind ChatGPT, in the middle of programming a physical robot, with a non-expert reviewing every program before it executes. The AI proposes the chain; the human decides what stays.
The system accepts the task in natural language, generates a candidate program, and renders it as visual blocks. Three things have to happen before the arm runs. The user has to read the plan, decide whether it matches their intent, and either approve, modify, or reject it. The explicit confirmation is the part the paper treats as load-bearing. The authors argue that an approval step is what turns an LLM from a black box into something a non-expert can actually direct.
The value compounds in the next step. Approved programs are saved to a personal library, and the system reuses those saved pieces on the next task. A block that has been trusted to sort plastic becomes a ready-made component the next time someone needs to sort plastic. Over weeks and months, the library grows into a working vocabulary for that robot, that workspace, and that user. The framework is built around three named goals: user trust, intent alignment, and reusability. The first two sit at the moment of approval. The third is what the approvals add up to.
This is the mechanism worth naming. APPROVE works, in effect, like a slow compiler, with the library standing in for the trusted codebase. A robot arm in a small workshop ends up with a set of trusted movements, the way a kitchen ends up with a set of trusted recipes. The library is the real payoff, and the approval step is how it gets built.
The framework has to clear two harder questions. The paper's mechanism assumes that approving a visual block is enough to understand what the robot will do in an unexpected situation, and that assumption is doing a lot of work. A block that says "check label" hides the camera work, the threshold for what counts as a label, and the failure mode when the label is missing. End-user programming research has a long history of prototypes that looked usable in a video and never escaped the lab, in part because the gap between reviewing a plan and owning its consequences turns out to be wide.
LLM-generated code adds a second test. It can run, look reasonable on screen, and still fail at the physical edge case the user never thought to test. A block that sorts plastic confidently in daylight can misroute in shadow, or double-pick a tangled pair of bottles. The approval step catches some of these, but only the ones the reviewer knows to imagine.
Whether the framework has a real future will turn on two things: whether the library pays for the cost of building it, and who gets to curate it when the stakes move from sorting recycling to something the user would rather not approve by eye alone.