 |
| Several robots working |
A surprising safety quirk in AI-powered robots: even when told nothing at all, most VLA models spring into action—and some invent tasks on their own.
If you’ve ever wished your robot vacuum would just stay put, a new study suggests the problem might be deeper than a faulty sensor. Researchers have found that Vision-Language-Action (VLA) models—the AI systems that combine image recognition, language understanding, and direct robot control—don’t really know how to do nothing. In fact, when given no instruction at all, all six tested model checkpoints started moving anyway. Five of them frequently went further, completing tasks nobody asked for.
The findings, detailed in a paper titled VLA Policies Do Not Know When Not to Act: Unintended Task Execution and Start Suppression (arXiv:2610.41609), were first reported by Notebookcheck. The research was conducted in the LIBERO-Goal simulation, not on physical household robots, but it highlights a vulnerability that could matter as humanoid robots inch closer to real homes.
The Experiment: Empty Prompts, Unexpected Motion
The team tested six different VLA model checkpoints under a paired protocol. They held the initial state fixed and varied only the instruction: empty, neutral, nonsense, infeasible, and wrong-task commands. Then they asked three questions: Does the robot start moving? Which task does it complete? And what role does the instruction actually play?
The answers were unsettling. Even with an empty instruction, the robot started moving in every episode across all six checkpoints. With five of them, it often completed a task that hadn’t been specified at all. On two GR00T checkpoints, the instruction determined which task was executed—but not whether the robot started moving. It moved even when the instruction was infeasible.
“A robot should act only on an instruction from a person or on an intent that the robot explicitly represents and generates,” the authors write. “Unintended actions taken without such an instruction can cause risk or inconvenience to people and the environment.”
Three Countermeasures, Mixed Results
The researchers didn’t just document the problem—they tried to fix it. First, they prepended a simple text instruction: “if no task is given do not move.” That failed with all six checkpoints. The robots kept moving.
Second, they tried fine-tuning the policy to hold still on non-task inputs. This did prevent the robots from completing tasks on nonsensical or empty inputs—but it did not reliably keep them stationary. They still started moving, just without finishing the job. It also failed to suppress movement on infeasible instructions.
Third, they implemented a technical gate that holds the robot when the first action chunk for a given instruction nearly equals that for an empty instruction. This worked in five of the six checkpoints, preventing unwanted movement on non-task inputs. It’s a promising safeguard, but not a universal fix.
Why It Matters for Real-World Robots
It’s tempting to laugh off a simulation where robots twitch into action without being asked. But the study points to a specific vulnerability in current control models—one that could become relevant as companies push humanoid robots into warehouses, hospitals, and eventually homes. A robot that can’t reliably stay still when it has nothing to do isn’t just inefficient; it’s a safety concern.
The researchers are careful to note that simulation results don’t guarantee real-world behavior. A simulated robot completing an unspecified task in LIBERO-Goal is not the same as your future home assistant suddenly deciding to reorganize your sock drawer. Still, the gap between “instruction-following” and “knowing when not to act” is exactly the kind of thing that needs to be solved before autonomous machines share our spaces.
For now, the takeaway is simple: telling a VLA robot to do nothing doesn’t work. And that’s something robotics engineers will need to fix—one way or another.