DynaHarness, a robot-control system built around a frozen action policy, succeeded in 75.2% of 800 newly sampled LIBERO-Pro simulation scenes, researchers report. The low-level policy alone succeeded in 17.5% of those same test episodes. The new scenes changed instructions or object positions within the benchmark’s existing task families; they did not test new task families.
The frozen policy is the model that turns camera input and a task request into robot movements. DynaHarness adds a vision-language planner that calls reusable skills, such as grasping or object recovery, then checks each command against the observed scene and its prerequisites before execution. If a command cannot be grounded or its requirements cannot be resolved, the system refuses it rather than send an unverified action to the robot.
During development, the system records where attempts fail so researchers can target changes to a skill or the execution logic. They test proposed and existing versions on the same task-and-seed pairs before keeping an update. Unlike SafeHarness, a separate robot study previously covered by GTN that checked planned routes for obstacle collisions, DynaHarness uses failure records to guide skill or execution changes and tests those changes across instruction and object-position perturbations.
The 75.2% figure is for the full DynaHarness stack against the frozen low-level policy, so it does not isolate the contribution of any single added component. On physical hardware, the paper reports 9 successes in 10 circle pick-and-place trials, 8 in 10 cup-stacking trials, and 7 in 10 trials each for bread pick-and-place and drawer placement. These four small task demonstrations are separate from the 800 simulation episodes; the authors say new task families and physical settings need their own evaluation.
Sources
- DynaHarness paper (arXiv) and full text.
- GTN’s earlier report on the separate SafeHarness collision-checking study.
- Illustrative lead image: robot arm handling an assay plate, courtesy of the U.S. National Institute of Allergy and Infectious Diseases (NIAID), public domain; not an image of the DynaHarness system.
