Skip to main content
Back to timeline
arXivSource publication:

An excavator stops taking language orders at face value: a one-use lease ties approval to the exact motion, cutting closed-site dumping from 1.23 to 0.33 cubic meters

Synopsis

An LLM-directed excavator can execute mechanically feasible motion whose destination or authority no longer matches the operator's approval. The work binds authorization to the exact motion: a one-use lease ties the action, generated motion, mission contract and admission tolerance together, atomic dispatch rechecks current authority and measured joint drift, and revocation requests a carry-pitch quintic stop. Across a 672-case interface suite the gateway admitted all 120 benign cases and produced zero unauthorized dispatches in the attack suite; in synchronous simulation, dump-phase revocation cut maximum closed-site deposition from 1.23 to 0.33 cubic meters.

Source-provided article image: From LLM Plans to Authorized Motion: Execution-Bound Authorization for Excavators
Fig. 1 · arXiv

Interpretation

Language-directed excavator motion now passes a check bound to that specific motion, so a changed authority or a substituted action after approval is rejected. Earlier checks either validated fully at approval time and only verified token authenticity at dispatch, or checked schema and mission limits, so post-validation revocation, target substitution and state changes still got through. In an interface suite of 552 attack and 120 benign cases, the complete method produced zero unauthorized dispatches and admitted every benign case, while schema-and-limit checking produced 240 unauthorized dispatches and validate-then-token checking produced 120.

When authority is revoked mid-execution, the machine generates a checked stopping trajectory instead of freezing the reference. Freezing the reference creates a substantial swing-drive transient in loaded slewing, and the stop itself releases material. In synchronous simulated dump-phase revocation, carry-pitch curl-back reduced soil released during the stop from 0.29 to 0.12 cubic meters and maximum closed-site deposition from 0.51 to 0.33 cubic meters; slew-phase stops used 74% of nominal swing torque, settled within 2.1 seconds and limited back-swing to 0.05 degrees.

Authorization splits into an admission tolerance at validation and an atomic recheck at dispatch, allowing measured joint drift inside that tolerance. Requiring exact state equality at dispatch rejects every drifted dispatch, while verifying only token authenticity lets post-validation revocation and substitution through. Across four loaded digging cases, state drift reached 0.0033 radians and stayed inside the admission tube, and all 52 delayed dispatches had nonzero drift; the exact-equality configuration therefore rejected 24 benign cases whose 4 mm drift was inside the 10 mm tube.

Perspective

The demonstrated scope is a 30.2-tonne excavator model: a 672-case interface suite, four loaded synchronous simulated digging cases, unloaded MuJoCo and Pinocchio transfers, and two local command-level planners. The loaded study uses a stationary work base, and closing a destination withdraws permission to deposit material there. It suits researchers and engineers working on authorization and runtime supervision for language-directed robots.

Loaded interruption uses synchronous switching: the plant is paused during stop synthesis and the generated reference starts from the reference state at detection, so the recorded stopping trajectories come from a paused plant; supervision intervals and motion durations use simulation time, while synthesis cost is reported separately in wall-clock time. Feasibility of the quintic stop depends on the duration search returning a sampled-admissible candidate, and the detection and installation delay for authority loss is set by the supervision period plus synthesis time.