A harmless chat line can move a commercial robot: TARS verifies cross-model action misalignment on four platforms
Synopsis
Language-driven commercial robots hand large-model reasoning to small action models for physical execution, leaving room for a semantic break between user intent and the motion actually performed. TARS treats that break as a black-box fuzzing problem, pairing action-aware and action-blind semantic search tracks with multimodal action observation and independent replay validation to find and reproduce unrequested physical actions on four commercial robots: AgiBot X2, Unitree G1, Unitree Go2 and Booster K1. On Unitree G1, TARS reaches up to 94.1% verified action coverage, showing that much of a robot's behavior can be reached indirectly without any explicit action command.
Interpretation
Commercial embodied robots can execute physical actions even when the user input contains no action command at all, a phenomenon arising from semantic inconsistency across large-model understanding, model-to-model collaboration and physical execution. The usual assumption is that an attacker must first issue an unsafe command to make a robot act unsafely; the observed behavior shows that seemingly benign conversation, ambiguous expressions or action-unrelated inputs can also trigger walking, turning, dancing or gesturing. Unitree Go2 executes a dangerous front jump without the required confirmation, Booster K1 can be induced to perform attacker-controlled multi-action sequences, and Unitree G1 turns a code-generation request containing straight_punch into a real punching motion.
TARS treats the language-to-action boundary as a fuzzing target, using two complementary semantic search tracks to automatically find prompts that trigger physical actions. Conventional fuzzing relies on crashes or internal program coverage as feedback, whereas commercial robots expose only black-box interfaces, the language-to-action mapping stays inside proprietary components, and large-model decisions are stochastic. The action-aware track builds indirect cues around known actions, the action-blind track explores semantically diverse inputs without revealing the target action to the generator, and both are driven by observable action-level feedback and filtered through multimodal action observation, default-behavior calibration and independent replay validation.
Across four commercial robots, 44 experimental settings and more than 130,000 executions, TARS consistently finds reproducible action-triggering prompts. Platform reactivity differs widely: Unitree Go2 shows raw trigger rates of 99.7% to 99.9% while Unitree G1 peaks at 43.3%, yet TARS reaches 100.0% action coverage on both platforms. With the Gemini 3.1 Pro backend, TARS reaches 94.1% verified action coverage on Unitree G1, verified triggers on AgiBot X2 reproduce with up to 100% success, and the triggers remain effective across independent sessions, contextual changes and natural prompt perturbations.
Perspective
Experiments ran on four commercial robots owned by the authors, all under direct supervision in a controlled laboratory, covering AgiBot X2, Unitree G1, Unitree Go2 and Booster K1 across 44 experimental settings; Unitree Go2 was excluded from the primary anonymous-track comparison because its default interaction policy produces a behavior for nearly every input.
Verified action coverage and trigger reproducibility diverge: Unitree G1 reaches up to about 94% verified action coverage, but its verified-seed reproducibility ranges from 28% to 90% across prompt-generation models, and Unitree Go2 is the least stable at 4% to 50%. Hallucinated intent completion accounts for roughly 89.3% of all observed divergence, showing that agreement between the large and small models does not imply alignment with the user.
