Robots that can finish a task are not the same as robots that can finish it the way you asked.
Researchers have built a vision-language-action (VLA) system called ExecVLA that teaches robots to care about how a job gets done, not just whether it gets done. Most VLA models are trained to hit a goal state, such as stacking a set of blocks, and treat every path to that goal as equally good. ExecVLA instead splits a robot's internal representation into two layers: one that tracks the invariant goal, and one that tracks execution details like which object to interact with, the motion pattern, spatial relationships, and the final position. The team trained and tested the approach on the LIBERO simulation benchmark and a real Realman-75 robot arm, using new datasets they built with both goal and step-by-step reasoning annotations.
This matters because real-world instructions are rarely just "do the task." They come with constraints: grab the left handle, not the right one; approach from above; leave the object upright when done. A robot that hits the broad goal while ignoring those constraints is a robot that is unreliable, or unsafe, in a warehouse, kitchen, or lab setting. ExecVLA's results show standard goal-completion scores improve, but the bigger gain is in actually following those fine-grained instructions.
It's a narrow, incremental fix, not a breakthrough; the testing is still mostly simulation with a single real robot arm. But it names a real blind spot in how these systems get graded: finishing a task and following orders are not the same thing, and most benchmarks have been rewarding the former.