Notre Dame's SAFI drops VLA prohibited-task success to 1.3%
TL;DR
- SAFI reduces OpenVLA-OFT prohibited-task completion after mid-execution takeover from the base policy's 99.7% to 1.3%, with no recorded physical incidents.
- The two-stage training labels prohibited demonstrations with stationary arm and gripper-hold targets, then trains jointly with retained-task actions using LoRA.
- Ablation shows removing retained-task loss collapses permitted-task performance to zero; removing fallback supervision lets the robot finish the prohibited task.
The original OpenVLA-OFT policy completes a prohibited task 99.7% of the time. A training recipe from University of Notre Dame researchers brings that rate to 1.3%, with no recorded physical incidents and no action-saturation episodes.
The paper, Teaching Vision-Language-Action Models to Stop Safely, posted on November 20, introduces State-Aware Fallback Imitation, or SAFI. Authors Chuhan Wang, Jinduo Guo, Yue Huang and Meng Jiang at Notre Dame, with Xiangchi Yuan at Georgia Tech and Ningshan Ma at MIT, target a failure mode that initial-state refusal benchmarks miss. "Policies which suppress a task from its initial state can still complete it when taking over partway through execution," they write. Stopping the task is not the whole problem either. "The robot may drop a held object, making non-completion an incomplete measure of safety."
SAFI's fix runs in two stages. Fallback target construction relabels every state in a prohibited-task demonstration with a stationary arm command and a gripper-preserving target, hold what you were holding. Joint policy adaptation then trains the model on those holding targets together with the original actions of retained tasks, using LoRA on the language-model layers and updating the inherited action head.
On LIBERO after mid-execution takeover, NPO-RT records 31.9% Target Success and a 19.9% Physical Incident Rate. GA and GradDiff push target rates further down but hit the controller's clipping bounds. SAFI "reduces target success from the original policy's 99.7% to 1.3%, with no recorded physical incidents or action-saturation episodes." On the BadLIBERO poisoned-object setting, which tests whether a backdoor-implanted policy can be stopped, SAFI "records zero on all three endpoints, matching RMU." Retained-task performance stays at 96.9% on the Spatial suite and 93.1% on Object.
The ablation is where the two components earn their place. Strip the retained-task loss and permitted-task performance collapses to zero. Strip the fallback supervision and the prohibited task gets completed again, at 99.7% on Spatial and 100.0% on Object, with physical incidents returning. Code is posted at franciscoliu/SAFI. The work lands in a 218-story run of robotics coverage we've tracked across the last 90 days, several of them probing exactly this gap between what a VLA refuses on paper and what it does once its gripper is moving.
Originally reported by huggingface.co
Read the original article →Original headline: Notre Dame SAFI Teaches VLA Robots to Abort Prohibited Tasks Mid-Execution Without Dropping Objects