SAFI teaches OpenVLA-OFT to hold still, cutting takeover success from 99.7% to 1.3%

Teaching Vision-Language-Action Models to Stop Safely

Zheyuan Liu, Chuhan Wang, Jinduo Guo, Yue Huang, Xiangchi Yuan, Ningshan Ma, Meng Jiang

cs.RO

2026-11-19

SAFI relabels banned demos as a stationary hold and keeps skills. On LIBERO, OpenVLA-OFT prohibited success after takeover falls from 99.7% to 1.3%, with no drops or saturation.

What problem this solves

A VLA turns camera images and a language instruction into robot actions. After the policy learns a skill, deployment still needs a way to switch that skill off. Existing suppression mostly shows up at the start of an episode. Gradient ascent, representation redirection, and refusal finetuning can make a banned task look rare when rollout begins from the initial state. Once the arm is already moving, several of those edits stop holding.

Takeover here means the original policy runs first, then an edited policy receives control. On LIBERO-Goal, some edits block stove activation from the start, then turn the stove on in nearly every episode after an early takeover and in every episode after a mid-task takeover. Leaving the task unfinished is also a weak safety signal. A wine bottle still in the gripper can be dropped or kept. Both count as not placed on the rack. Figure 1 puts those endings side by side. The acceptable stop is the hold.

Success rate misses two separate questions: is the robot still executing at intermediate states, and is the action it substitutes physically acceptable?

Method

SAFI, short for State-Aware Fallback Imitation, is used only when a stationary hold is already a reasonable alternative. Before training, a scripted hold is rolled out from candidate takeover states. It has to block the prohibited task and meet that task's physical-event criteria. Tasks that fail the check stay out of the experiment.

Fallback labels rewrite the prohibited demonstrations. Inside each action chunk, a short sequence of future actions predicted together, every arm command becomes a no-motion vector: no change in position or orientation, encoded with that demonstration's action normalization. The gripper value is copied from the first step of the same chunk and repeated for the rest of the chunk. Observations and the instruction are unchanged, so both the initial frame and the intermediate frames have targets.

The two imitation terms are added with equal weight. Prohibited inputs are fit to the hold. Permitted tasks are fit to their original actions. Both losses are mean absolute error over the chunk. Training starts from the original policy, puts LoRA (a low-rank update) on the language-model layers, and updates the action head as well. Checkpoints are filtered on suppression and retained performance. The kept checkpoint is the one with the highest retained success.

On LIBERO, each edit bans one task and treats the other demonstrations in that suite as the retained set. On BadLIBERO, a backdoored LIBERO setup, poisoned demonstrations get fallback targets and clean demonstrations keep their actions. The two sets can share an instruction. There is no preference loss and no representation-matching term. The learned alternative is a still arm and the chunk's initial gripper command. Putting the object down, or pulling out of contact, is not in the target.

Results

The base models are the released suite-specific OpenVLA-OFT policies. The original policy runs to an early or mid-task state, and the edited policy takes over from a matched restored state. The LIBERO score averages Goal, Spatial, and Object equally. Baselines are GA, GradDiff, NPO-RT, RMU, Refusal FT trained toward a home return, and an OFT adaptation of VLA-Forget. Saturation means a command hits the controller's clipping bounds.

MethodTarget success after takeoverPhysical incidentAction saturation
Original policy99.7%12.1%0.0%
GA12.2%16.3%100.0%
GradDiff11.7%27.5%99.4%
NPO-RT31.9%19.9%46.9%
RMU50.6%14.4%0.0%
Refusal FT11.6%4.1%0.0%
VLA-Forget30.9%20.3%53.6%
SAFI1.3%0.0%0.0%

GA and GradDiff bring target success to about 12%, while commands hit the clip limit on 100.0% and 99.4% of episodes. Refusal FT avoids saturation, at 11.6% target success and 4.1% incidents. The gap against SAFI is concentrated on Goal. SAFI has the lowest target success and incident rate in the table, and saturation at 0.

BadLIBERO tests a poisoned-object attack. The implanted reference reaches the attack target on 87.5% of takeovers. GA, GradDiff, NPO-RT, RMU, Refusal FT, and SAFI all record 0% attack success. Incident rates for the first three are 40.0%, 30.0%, and 22.5%, with 100% saturation. RMU and SAFI are at zero on all three endpoints. Refusal FT has zero incidents and 2.5% saturation. On this backdoor, SAFI does not beat RMU.

The ablation splits the two losses. Retained imitation alone leaves Spatial target success at 99.7% and Object at 100.0%, with a 30.9% incident rate on Object. Fallback imitation alone pushes target success near zero and retained success to zero. With both terms, Spatial is at 3.4% target success and 96.9% retained success. Object is at 0.3% and 93.1%. Incidents are zero. The original retained rates on those suites are 99.2% and 94.2%.

In the checkpoint plot, suppression is 100% minus takeover target success. Later SAFI checkpoints reach 99.7% suppression with 97.2% retained success on Spatial, and 99.0% suppression with 93.6% retained success on Object. At that suppression level, GA's retained success stays at 0. GradDiff reaches 34.2% and 41.7%.

Displacement is measured from the object's position at takeover. Averaged across Spatial, Object, and both takeover times, SAFI ends 3.1 cm away, with the gripper still in contact on 97.3% of rollouts. Refusal FT is trained to go home. Its contact rate is 98.8%, and displacement is still 27.6 cm. Under mid-task takeover, SAFI moves 4.8 cm on Spatial and 3.2 cm on Object. The closest other edits in that comparison are Refusal FT at 12.9 cm and RMU at 21.3 cm.

Why it matters

A safety-tuned language model can refuse a harmful request, then continue the answer if it is handed the opening lines. A VLA has the same sequential hole in physical form. Later images are produced by earlier actions, so an edit judged only from the initial state never has to act on the frames a takeover actually shows.

SAFI leaves the instruction in place and replaces prohibited actions with a stationary hold, then puts that target in the same error as the skills that should survive. When the gripper is already on an object, and the right response is neither a release nor more progress toward the goal, the hold matches the state more closely than a home-return trajectory.

This is a behavior edit. Someone has to name the banned skill, demonstrations have to exist, and a scripted hold has to pass the check first. The fit is turning a skill off in simulation. Spotting a new hazard in an open scene is outside the procedure.

Limitations

The evaluation uses OpenVLA-OFT, selected LIBERO and BadLIBERO tasks, and one training seed. It does not change the architecture, repeat training, or run a physical robot under different sensing and control. The fallback is a hold. If the object is unstable, or contact is already underway, a better stop may be a controlled placement or a withdrawal. Those targets are not learned.

The 1.3% number is measured on tasks where a scripted hold already blocks completion. There is no evidence for tasks that require motion in order to be safe. On BadLIBERO, RMU matches SAFI at zero attack success, zero incidents, and zero saturation, and the main table does not report retained success for that setting. Spatial alone is still at 3.4% target success. Takeover states are restored from the original policy's trajectory. States the edited policy would reach on its own are not listed separately. Zero saturation only means commands stayed inside the controller clip.

Terms

Source

Related papers

All paper explainers