Integrating Large Language Models (LLMs) into heavy machinery like excavators promises a future of intuitive, natural-language task planning. However, this "conversational" interface introduces a massive attack surface—from prompt injections that could redirect a 30-ton machine to dump material in a restricted zone, to the simple reality that site conditions (and thus authorization) can change in the seconds between a command being issued and the machine completing its swing.
This paper from researchers at Tampere University, Korea University, and Sookmyung Women’s University introduces a framework for execution-bound authorization, ensuring that every inch of a robot's motion remains strictly within the bounds of current operator approval.
TL;DR
The researchers developed a security framework that binds high-level LLM plans to specific, pre-certified physical trajectories for excavators. By using one-use leases and 10ms supervision loops, the system can detect revoked permissions mid-motion and execute a "carry-pitch" stop that safely halts the machine while preventing unauthorized material deposition.
What Problem Does This Paper Address?
Robotics research is rapidly adopting LLMs for task planning, but these models are notoriously susceptible to "jailbreaking" and prompt injection. In an industrial setting, an attacker could trick an LLM-directed excavator into performing a mechanically feasible but unauthorized action—like dumping soil onto a worker or a closed site.
Furthermore, industrial environments are dynamic. As shown in Figure 1, an operator might close a dump site after a skill has already been dispatched. Traditional "check-at-start" authorization fails here. The challenge is bridging the gap between high-level intent and low-level hydraulic execution while accounting for "Time-of-Check-to-Time-of-Use" (TOCTOU) races and physical drift.
Methods
The authors propose a three-stage "approval–dispatch–execution" interval protection:
- Motion-Bound Leases: When a plan is proposed, the gateway generates a one-use lease. This lease cryptographically binds the high-level action to a specific motion descriptor (the exact joint trajectories), the current mission contract, and a calculated admission tolerance.
- Atomic Dispatch: Right before execution, the system re-verifies authority and checks if the machine's physical state has drifted. They use a first-order kinematic allowance ($\rho$) to ensure the machine is actually where the plan thinks it is.
- Periodic Supervision & Quintic Stopping: A monitor checks authority every 10ms. If permission is revoked (e.g., a site is closed), the system triggers a quintic stop. Unlike a hard emergency stop, this is a smooth, jerk-minimized trajectory that includes a "carry-pitch" endpoint—curling the bucket back to retain the load during the stop.
The workflow is visualized in Figure 3, showing how a revocation at time $t_r$ triggers a preemptive stop before the machine can complete an unauthorized dump.
Datasets
The framework was evaluated using several rigorous benchmarks:
- Interface Suite: A custom suite of 672 cases, including 552 "attack" cases (prompt injection, replayed approvals, stale evidence) and 120 benign cases.
- Loaded Digging Scenarios: Four distinct operating cases in a high-fidelity synchronous simulation (MuJoCo/Pinocchio) varying soil density, digging depth, and friction.
- LLM Planners: The system was tested against plans generated by Qwen2.5-0.5B and 3B models to see if it could catch "jailbroken" or incorrect proposals at the gateway.
Key Insights and Ideas
- Binding Identity to Motion: You cannot just authorize the intent (e.g., "dump soil"); you must authorize the exact trajectory. This prevents an adversary from substituting a safe-looking command with a dangerous physical motion.
- The "Carry-Pitch" Stop: Safety in excavation isn't just about stopping the metal; it's about controlling the material. By curling the bucket during a stop, the system significantly reduces "spill" into unauthorized zones.
- Handling Drift: The "admission tolerance" is a clever mathematical way to handle the reality that machines move slightly due to gravity or hydraulics even when "holding." If the machine drifts too far from the plan's starting point, the lease is invalidated.
Key Results
The results demonstrate that the framework is both secure and physically performant:
- Security: The gateway achieved zero unauthorized dispatches across the entire 552-case attack suite, while correctly admitting all 120 benign cases.
- Material Retention: During a dump-phase revocation, the framework reduced the amount of soil deposited in a closed site from 1.23 $m^3$ to just 0.33 $m^3$.
- Mechanical Integrity: The quintic stop was much gentler on the machine than a simple "freeze" command. It used less than 75% of nominal swing torque, whereas a sudden freeze caused a 102% torque spike and significant tracking errors.
- Latency: Gateway checks were extremely fast, with a median call time of just 0.50 ms, making it suitable for real-time industrial control loops.
As seen in Figure 2, the "Execution-bound" system (frames 2 and 3) successfully holds the load when the site is closed and only completes the dump once the site is reopened, whereas a standard "Dispatch-only" gate (frame 1) fails.
Limitations and Future Work
- Stationary Base: The current study assumes the excavator's base is stationary. Extending this to mobile excavators would require more complex admission tolerances that account for base movement.
- Synthesis Time: While the checks are fast, generating the optimal stopping trajectory (synthesis) took a median of 142 ms. In the simulation, the "plant" was paused during this time. A real-world deployment would need to optimize this synthesis or use a pre-computed buffer of safe stops.
- Trusted Base Assumption: The framework assumes the gateway and executor themselves are not compromised. It protects the machine from external planner attacks, not from low-level firmware compromises.
Summary updated