Robot execution stops trusting finished actions: LOGIV tracks broken conditions on a task graph, lifting real-robot success from 12/30 to 17/30
Synopsis
In long-horizon robot tasks, an action can finish while the condition it established is later disturbed, yet the remaining plan still assumes that condition holds. LOGIV compiles a validated plan into a task graph with causal and resource-ordering edges, checks those dependencies during execution, and after a failure repairs only the affected part while retaining still-usable unfinished actions. Across 10 dual-arm tasks each from RoboTwin V2 and BiCoord and three execution backbones, mean success without injected disturbances rises from 71.67% to 82.33% and from 42.33% to 57.67%; on three real-robot tasks successful trials rise from 12/30 to 17/30.
Interpretation
The robot execution layer gains a continuous check: whether an action is complete is recorded separately from whether the condition it supplies to later actions still holds, so a broken condition need not reopen finished actions. Prior execution organization advanced on action completion, but a support established by a finished placement can be invalidated by a later disturbance while the remaining plan still assumes it. Across 10 dual-arm tasks each from RoboTwin V2 and BiCoord with the π0.5, AHA-WAM and Light-WAM backbones, mean task success without injected disturbances rises from 71.67% to 82.33% and from 42.33% to 57.67%.
Repair after a failure is confined to the affected actions, the other unfinished actions are kept as they are, and they constrain which conditions must be restored. The comparison is replanning all remaining actions, or choosing repair nodes by their position in the linear plan. Removing causal dependency tracing and selecting nodes by linear position keeps the modified-node fraction at 24.6% but drops clean-task success from 72% to 61% and perturbed-task success from 60% to 52%; replanning all remaining actions raises repeated valid actions from 0.2 to 1.6 and lowers perturbed-task success to 54%.
The repair fragment and the retained actions are joined into a complete plan that is checked against the symbolic model before execution resumes, rather than accepting the local fragment alone. A local repair fragment generated without checking the combined continuation can conflict with the retained actions. Without logical validation of the repaired plan, condition restoration falls from 72% to 62% and perturbed-task success from 60% to 57%.
Perspective
The evaluation covers 10 dual-arm tasks each from RoboTwin V2 and BiCoord with three execution backbones, plus three real-robot tasks on Astribot S1 and AgileX Piper platforms with 10 trials per task; no disturbances were injected in the real-robot trials, and the PDDL specification is predefined and held fixed.
Real-robot trials are 10 per task, and the authors note wide confidence intervals, so the result indicates a trend rather than a precise effect size; most residual failures in the cable-insertion task occur when the plug is aligned but fails to seat, depending on the low-level policy's contact-rich precision; repair terminates when no further expansion is possible or the budget is exhausted, without guarantees of minimality or completeness.
