DynaHarness: A Dynamic Physical Harness for Self-Evolving Robot Agents
DynaHarness is an arXiv preprint that adds a dynamic physical harness around a frozen vision-language-action policy: a slow Qwen3-VL-4B planner proposes capabilities, while a fast 2 Hz governor grounds, monitors, substitutes and recovers them. On LIBERO-Pro it reaches 75.2% success on 800 newly sampled states versus 17.5% for the frozen π0.5 policy, and its self-evolution gains transfer to new states. This gives robot-agent builders a retraining-free way to govern and improve existing policies from execution evidence.
How should robot agents govern physical execution and convert failure evidence into validated, reusable revisions of capabilities and execution mechanisms?
Long-horizon manipulation requires tight coupling between slow semantic reasoning and fast physical control, but existing harnesses either rely on coarse high-level replanning or improve a fixed skill library without identifying which component caused a failed episode. Episode-level failure gives little guidance for revision, and running commands lack explicit authority over grounding, refusal, budgets and termination.
Prior work falls into two families: execution-orchestration harnesses such as Harness VLA, PhyAgentOS, EmbodiedSkills, ETA/Show-Harness and Thea decompose tasks and add verification or recovery, but they do not link execution evidence back to targeted revisions. Self-improvement methods such as ASPIRE, GaP, InSight, ENPIRE and Zetta repair programs or acquire skills, but task-level failures may originate in planning, grounding, execution, verification or recovery, and they lack shared command-level provenance and paired admission checks.
DynaHarness uses a shared physical execution contract: every robot-facing command is bounded by budget and lease, checked at 50 Hz, and recorded. A slow brain, Qwen3-VL-4B, is called on demand and proposes only a capability name and symbolic arguments; a deterministic fast brain at 2 Hz binds parameters, checks preconditions, refuses unresolved actions, substitutes capabilities, invokes recovery and requests replans while a 20 Hz controller executes. Analytic contact skills, recovery skills and a frozen π0.5 VLA share the same contract. Offline, attributable episode records pass through ordered diagnostic layers; failures direct physical-quantity revisions, and paired regression checks plus broader validation admit or reject them.
On LIBERO-Pro development seeds, DynaHarness with Qwen3-VL-4B scores 74.25% averaged over four cells, versus 16.2% for frozen π0.5 and 20.3% for PhyAgentOS with the same upper model. On 800 newly sampled states after freezing, it scores 75.2% versus 17.5% for π0.5, with 473 versus 11 exclusive wins; four snapshots preserve rank from development to new states, with a 14.4-point transferred gain versus 14.3 during development. On 261 untouched official states it reaches 77.0% versus 16.5% for the policy. In paired ablations, Full scores 74.0%; disabling failure-triggered replanning, substitution and recovery reduces success to 63.9% (A2static), and freezing a skill sequence gives 63.8%. Removing seven analytic contact skills drops success to 16.6%, near bare π0.5 at 16.2%; removing recovery/intervention only lowers success to 73.1%. Self-evolution raised development success from 60.0% to 74.25%; a drawer-related update improved one goal-swap cell from 55% to 90–100% over 20 seeds and Goal success from 75.25% to 78.0% over 400 episodes, but a candidate that passed its paired gate lost 79 successes on five stove-hob cells and was reverted. Real-robot trials reached 90% on circle pick-and-place, 80% on cup stacking, 70% on bread pick-and-place and 70% on drawer placement, with 10 trials per task.
The paper does not include a separate limitations section. The main evaluation is on LIBERO-Pro simulation, and real-world testing is limited to four tasks with 10 trials each, so broader hardware generality is not established. The system relies on the benchmark success predicate for monitoring and on task-specific analytic and recovery skills; 10-S remains the weakest suite at 65.0% development and 66.5% on new states. Evolution uses the development seeds and cells that motivated revisions, and a candidate that passed the paired gate still caused losses elsewhere until broader validation rejected it. No code release is mentioned, and compute/latency costs of validation and slow-brain calls are only partially reported.
DynaHarness is most relevant to builders of manipulation stacks on frozen VLA policies, such as robot-arm OEMs, warehouse and logistics automation providers, and service-robot software vendors. It could act as a runtime governance and maintenance layer that refuses unsafe commands, substitutes capabilities and records evidence for field updates without retraining the VLA. In the near term, simulation and benchmark teams can use it for policy evaluation and repair; real deployments within 1–3 years would require a reliable success or verification interface, safety certification for automated library changes, and evidence that development gains transfer across robots and environments. Over 3+ years, it could support continuous self-improvement of deployed fleets if paired validation becomes robust and affordable.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.