VISTA: Active Exploration for Self-Improvement in Embodied Agents

• Long-horizon manipulation is difficult for end-to-end policies amid compounding errors, perturbation shifts, sample demands, and long-tail actions across task phases.Source: §1 · pp. 1–2
• Existing execution records may lack observations or interventions needed to test failure explanations, motivating direct environmental evidence for harness improvement.Source: §1 · p. 2
• VISTA treats the operational harness as an evolvable executable control plane, linking active diagnostic exploration to revisions and preserving lineage for hypothesis retrieval and mechanism reuse.Source: §1 · pp. 2–3
• Across LIBERO-Pro, RoboCasa, and RoboTwin, VISTA progressively improves manipulation success over the corresponding initial harnesses.Source: §1 · pp. 2–3

2.1 Foundation Models for Robotic Manipulation

• Low-level policies provide physical control but are embodiment-tied, while semantic representations preserve generalization but depend on grounding pipelines and can limit physical adaptability.Source: §2.1 · p. 3

2.2 Embodied Agent Systems

• Agentic embodied systems close the perception–decision–execution loop through harnesses that use programs, skills, learned policies, or planners, alongside memory, feedback, and recovery.Source: §2.2 · p. 3
• Callable skills, controllers, and VLAs simplify planning, but opaque executors reduce the model’s role in fine-grained physical control.Source: §2.2 · pp. 3–4
• Prior work improves harnesses, skills, critics, and recovery through optimization or exploration; VISTA links meta-controlled exploration to executable revisions and preserves code, evidence, and lineage.Source: §2.2 · p. 4

3 Method

• VISTA keeps model parameters fixed while a meta-agent develops harness candidates through active investigation, revision, and autonomous testing.Submitted candidates are frozen and evaluated in fresh autonomous episodes; their versions and evidence remain available for later iterations.Source: §3 · p. 4, §3.1 · p. 5

3.1 Harness DAG

• Harnesses package instructions, executable skills, policy hooks, persistent knowledge, and tool bindings; immutable Git-versioned versions form a DAG retaining alternatives and evidence.The meta-agent can inspect earlier implementations and reuse mechanisms, while unsuccessful proposals preserve diagnostic findings for later development.Source: §3.1 · p. 5

3.2 Evidence-Guided Harness Revision

• Aligned records of planner decisions, tool calls, robot actions, images, and state help localize failures and guide revisions to implicated harness components.When existing trajectories cannot resolve a hypothesis, VISTA acquires additional embodied evidence.Source: §3.2 · p. 5

3.3 Active Diagnostic Exploration

• Before submitting a candidate, the meta-agent can use a private simulator session to inspect a frozen harness state and test a specific hypothesis with public capabilities.These interactions provide diagnostic evidence for revisions rather than modifying formal evaluation.Source: §3.3 · p. 5

3.4 Autonomous Validation and Selection

• Fresh, intervention-free formal episodes evaluate frozen candidates on matching complete matrices; ties retain the incumbent, and optional early stopping requires consecutive perfect evaluations plus confirmation.The method bounds search by configured development and iteration budgets.Source: §3.4 · pp. 5–6

4.1 Experimental Setup

• The evaluation covers four LIBERO-Pro suites under instruction-redirection and position-swap perturbations, two RoboCasa365 categories, and 50 RoboTwin tasks.LIBERO-Pro and RoboCasa365 use eight scene seeds with two repetitions; RoboTwin uses four randomized scenes with two repetitions.Source: §4.1 · pp. 6–7

4.2 Main Results

• LIBERO-Pro success rises from 17.5% to 95.0%; RoboCasa365 success rises from 16.4% to 24.2% on Composite Seen and 4.7% to 13.3% on Composite Unseen; RoboTwin success rises from 48.5% to 74.0%.The LIBERO-Pro gain over H0 is 77.5 percentage points; evolved harnesses reach 97.0% on T settings and 93.4% on S settings.Source: §4.2 · p. 7, Table 1 · p. 7, Table 2 · p. 7
• On LIBERO-Pro, Spatial, Object, and Goal improve rapidly in early rounds, while LIBERO-10 benefits from a longer sequence of revisions.Figure 3 curves average selected-harness success over the ten tasks in each setting.Source: §4.2 · p. 7, Figure 3 · p. 8

4.3 Ablation

• After two rounds, active exploration gains 85.9 percentage points using 91.6M mean cumulative Meta tokens per task, versus 56.3 points using 137.7M for passive exploration.After one round, active gains 50.0 points using 42.2M tokens, versus 39.1 points using 94.2M for passive exploration.Source: §4.3 · p. 8, Table 4 · p. 9

4.4 Analysis: Inside Harness Evolution

• Harness development can proceed through linear refinement, branching, and branch recombination, while retaining the best-so-far version when an attempted change fails to improve it.On Goal-t08, a candidate reaches 100% formal success after alternative stopping-rule exploration is rejected and another branch adds a metric placement skill.Source: §4.4 · p. 8, Figure 4 · p. 9
• Developed harnesses accumulate observation-dependent targets, grasp and execution-state verification, bounded recovery, and programmatic action sequences alongside VLA control.On Goal-t09, revisions to lift verification and release behavior raise autonomous success from 0.0% initially, to 56.3%, then 100.0%.Source: §4.4 · p. 9, Figure 5 · p. 10

• VISTA evolves the operational harness while keeping the underlying model parameters fixed.Source: §5 · p. 10
• Across LIBERO-Pro, RoboCasa, and RoboTwin, task success progressively improves over successive development rounds.Source: §5 · p. 10
• The authors report that physical real-robot validation has not yet been performed; current evaluations are limited to high-fidelity simulation environments.Source: §5 · p. 10

• Historical inspection reads completed records without recreating the prior simulator state, while live development starts a private scene from a frozen harness snapshot.Source: §A.1 · p. 15
• Autonomous evaluation runs fresh episodes with a frozen candidate and no meta-agent intervention.Source: §A.1 · p. 15
• Historical depth is available only for trace modes that archive metric depth and calibration; full_rgb records contain RGB but no historical depth.Source: §A.1 · p. 15

• The evolve-harness skill guides history inspection, hypothesis formation, harness revision, live-session diagnosis, and validation; it does not change planner or VLA weights.Source: §A.2 · p. 15
• The online harness can expose file and termination tools, image reading, observation and VLA-step skills, plus optional policy hooks within the host planner loop.Source: §A.2 · p. 16
• Meta tools control development and inspection, whereas harness tools determine online-agent capabilities during autonomous episodes; a meta-agent live call does not establish that the online agent had that capability.Source: §A.2 · p. 16

A.3 Git-Backed Harness History

• Each frozen harness is stored as a Git commit, with evaluation requests, episode records, live diagnostics, and selection decisions attached in the run store.Ordinary revisions create parent edges, while explicit merges create commits with multiple parents.Source: §A.3 · p. 16
• Rejected candidates remain reachable through retained references and can later serve as development bases, separately from the incumbent used for formal comparison.Reusing a file or idea does not create a Git ancestry edge.Source: §A.3 · p. 16
• Git records code lineage, while the run store records the evidence and decisions associated with each harness version.Source: §A.3 · p. 16

B.1 Complete Per-Task Results

• Tables A2, A3, and A4 report per-task success rates, with each task evaluated on eight scene seeds and two repetitions per seed.H0 is the unmodified harness; Evolved is the retained champion.Source: §B.1 · p. 17
• VISTA improves over H0 on all 40 LIBERO-Pro LIBERO-T tasks, with 34 reaching 100.0% autonomous-evaluation success.Suite averages rise from 19.0% to 98.0% on Spatial, 49.0% to 100.0% on Object, 16.0% to 99.0% on Goal, and 7.0% to 89.0% on LIBERO-10; six tasks remain below 100%, including LIBERO-10 T4 at 63.0% and T9 at 44.0%.Source: §B.1 · p. 17
• Table A3 reports per-task success rates for the four LIBERO-S suites of LIBERO-Pro.Source: §B.1 · p. 17, Table A3 · p. 18
• RoboCasa Composite Seen success rises from 16.4% to 24.2%, while Composite Unseen rises from 4.7% to 13.3%.Eleven of 16 tasks improve on Composite Seen and 11 tasks improve on Composite Unseen; Table A4 reports per-task success rates.Source: §B.1 · p. 18, Table A4 · p. 19

B.2 Complete Evolution Details

• Figure A2 separates round-level candidate performance from best-so-far champion scores and shows retained-champion gains over H0.Heatmap cells report candidate successes out of 16 episodes; dashed cells carry forward the last retained result after development stops, and an em dash marks incomplete evaluation.Source: §B.2 · p. 19, Figure A2 · p. 20
• Across suites, aggregate gains are largest in early rounds; Object reaches perfect development performance, while Spatial and Goal approach saturation and LIBERO-10 improves more gradually.LIBERO-10 retains a larger performance gap at the endpoint, and task progress is asynchronous; Goal T1 reaches 100.0% only in R10.Source: §B.2 · p. 19
• Candidate scores can be non-monotonic even when the retained champion stays unchanged, so a plateau in aggregate champion scores can coexist with continued exploration.The non-decreasing champion curves reflect best-so-far selection, not improvement by every proposed revision.Source: §B.2 · p. 19