Back to articles

Teaching VLA Robots to Stop Safely Mid-Task

2 min read

Vision-language-action (VLA) models let robots act on visual observations and language instructions. But safe behavior is not just a matter of refusing a task before it begins. If an action becomes prohibited after execution has started, a robot must respond to the state it is already in—and manage any object it is holding.

The authors report that policies trained to suppress a task from its initial state can still complete it when a takeover occurs partway through execution. A robot may also fail to finish the task yet drop the object it was carrying. That makes task non-completion an incomplete safety measure: the physical response to stopping matters too.

The paper introduces State-Aware Fallback Imitation (SAFI) for tasks in which continuing to hold an object is an appropriate alternative. The method has two stages:

  • Construct fallback targets: In demonstrations of prohibited tasks, assign stationary commands to the arm and consistent gripper targets. This provides explicit guidance for what the robot should do at intermediate states, rather than merely labeling the original task as something not to complete.
  • Adapt the policy jointly: Train the fallback targets alongside the original actions for tasks that remain allowed. The aim is to add a stop response while preserving the robot’s other skills.

The team evaluates OpenVLA-OFT on LIBERO and BadLIBERO tasks, measuring prohibited-task completion, physical incidents, and performance on retained tasks. On the evaluated LIBERO tasks, SAFI reduces aggregate target success after takeover to 1.3%. The experiments record no physical incidents or action saturation, while retained-task success remains high. These results highlight a useful shift in how robot safety can be assessed: not only whether a prohibited goal is reached, but also what the robot does physically when interrupted.

SAFI makes stopping a behavior to be learned, rather than relying solely on suppression of the intended outcome. Its scope is nevertheless specific. The fallback of holding is suitable only where that response makes sense, and the reported results come from a particular model and benchmark tasks. They do not establish a general guarantee across robots, objects, or real-world settings. Further evaluation will need to test other fallback behaviors and a wider range of takeover states.

Hugging Face Daily Papers

Comments

Checking sign-in status...

Loading comments...

Related articles

CCTest · Blog
EgoX Enables Robots with Different Bodies to Share Manipulation Experience
Robotics & Physical AI
cctest.ai

EgoX Enables Robots with Different Bodies to Share Manipulation Experience

EgoX introduces a cross-embodiment framework that organizes demonstrations from humans, physical robots, and simulated agents through modular observation and action interfaces. Its embodiment-dreaming objective improves transfer to target robots, especially under out-of-distribution conditions.

Read more