A Safe Action Is Not Enough: Feasible-Future Decoding for Vision-Language-Action Policies

• A likely, locally admissible move can leave no policy-supported route to safe completion, and uncertain execution makes successor states depend on both the action and environment.Reducing cost by ending an episode early can leave the task unfinished.Source: §1 · p. 2
• The exact next-block marginal combines the action prior, local reward–cost terms, and weighted mass of safe, successful continuations.The feasible-future mass records whether safe continuations remain and how much weighted mass they retain.Source: §1 · pp. 2–3
• Because exact evaluation is generally impractical online, the selective approximation gates intervention before reranking a small candidate set using policy likelihood, local robustness, and bounded continuation scores.Its recovery conditions account for intervention, candidate availability, support handling, and approximate-scoring errors.Source: §1 · pp. 2–3
• Across six Safety-CHORES settings, VICS-G consistently lowers observed safety cost while keeping task completion and episode length close to policy sampling.Source: §1 · p. 2
• On Fetch, VICS-R uses safe-completion estimates under a limited simulation budget and improves observed safe success while further reducing cost.Execution feedback also yields a favorable observed safety–completion trade-off under perturbed actions, with a matched ablation supporting preservation of task success within the preset tolerance.Source: §1 · pp. 2–3

• The frozen policy, transition kernel, and observation process induce a history-conditioned law over future trajectories; the target set contains trajectories completing the instruction within the designated hard safety budget.Source: §2 · p. 3
• The decoder chooses from a finite policy-supported candidate set, while its local robustness score combines policy likelihood with the predicted next-block safety margin.The candidate set retains the policy sample alongside top-K actions, with size at most K + 1.Source: §2 · p. 3
• A feasibility–likelihood gap occurs when the policy favors one candidate even though another has higher probability of safe task completion under its own continuation process.That probability averages over latent-state uncertainty, environment outcomes, and subsequent frozen-policy actions.Source: §2 · p. 3
• An unsafe shortcut violates a local constraint, whereas a safe dead end passes the local check but leaves zero frozen-policy probability of safe completion.Another controller may still be able to complete safely from a dead end for this policy.Source: §2 · p. 3

3.1 A FEASIBLE-FUTURE IDENTITY

• The safe-completion reference law weights complete trajectories by reward and cost; marginalizing future continuations yields a next-block marginal combining policy prior, local terms, and feasible-future mass.The reference distribution is not a claim that the deployed decoder solves a global control problem.Source: §3.1 · p. 4
• Feasible-future mass integrates execution, latent states, future observations, and policy actions; its support marks positive safe-completion probability, while its magnitude grades weighted continuation mass.A zero mass is relative to the frozen continuation process, not proof that every controller would fail.Source: §3.1 · p. 4

3.2 SELECTIVE FINITE CANDIDATES: FOUR APPROXIMATION INTERFACES

• The recovery analysis separates intervention, coverage, support, and ranking losses, and recovers the best retained viable candidate only under support-separation and score-margin conditions.The bound applies only when reranking is invoked, the surviving set is nonempty, replacement is authorized, and the theorem assumptions hold.Source: §3.2 · pp. 5–6

3.3 VICS: SELECTIVE FEASIBLE-FUTURE RERANKING

• VICS-G selectively revises the policy sample using likelihood, local robustness, and a bounded continuation proxy based on monitor and recent-execution features.The proxy is not a calibrated estimate of log Zfeas and does not certify support.Source: §3.3 · pp. 6–7
• VICS-R explores simulator continuations from reconstructed states, but its continuation law and selection rule differ from the exact target.A learned safe-completion critic is left for future work.Source: §3.3 · p. 7
• VICS-S uses symbolic knowledge to constrain eligibility before generic scoring, but viable actions may be rejected and dead ends admitted.Symbolic eligibility is a partial view of support, not knowledge of which candidates truly permit safe completion.Source: §3.3 · p. 7

4 EXPERIMENTAL EVALUATION

• The evaluation compares decoders on shared task episodes and seeds, reporting success, cumulative cost, violations, and episode length; the six-setting results are descriptive.VICS-G uses one shared configuration across settings, selected using approximately 10% of episodes per task family; matched controls examine individual components.Source: §4 · p. 8, §4.1 · p. 8

4.2 MAIN RESULTS

• VICS-G lowers observed mean safety cost by 1.9%–57.5% across six settings, with fewer violations in every setting, success within 2.5 percentage points, and episode length within 0.82 steps of policy sampling.The reported settings span task-optimized and safety-aligned checkpoints; Table 1 presents descriptive aggregates.Source: §4.2 · p. 8, Table 1 · p. 9
• VICS-S shows task-specific gains, while RCD trades lower raw cost for shorter episodes and reduced success, and continuation-correction gains vary by setting.On base Fetch, VICS-S raises success by 3.5 points and lowers cost by 46.0%; RCD success falls by 3.5–20.3 percentage points and episodes shorten by 4.4–37.5 steps; at ηv = 0.10, the correction lowers cost in four of six settings but has an adverse ObjectNav-base case.Source: §4.2 · p. 8

4.3 EXECUTION FEEDBACK UNDER IMPERFECT EXECUTION

• Under perturbed safety-aligned Fetch execution, feedback-aware VICS-G reaches 55.67% success, with 15.9% lower observed cost and 16.6% fewer violations than policy sampling at virtually identical mean episode length.Against feedback-masked VICS-G, feedback has 9.7% lower observed cost and a +0.58-point success difference; the cost interval spans zero, leaving its incremental safety benefit uncertain.Source: §4.3 · pp. 9–10

4.4 LOOKING AHEAD TO PRESERVE SAFE COMPLETION

• On safety-aligned Fetch, VICS-R raises observed safe success over VICS-G from 27.91% to 30.23%, with task success unchanged at 58.72%.Mean cost falls from 3.233 to 3.116 (3.6%); VICS-R uses at most one early lookahead decision and is tested in a modified, branchable simulator.Source: §4.4 · p. 10

4.5 LIMITATIONS

• The evaluation covers three simulated task families, and broader disturbances and physical-robot deployment remain future work.Decoder gains may vary across tasks and checkpoints; theoretical bounds depend on score accuracy and candidate coverage.Source: §4.5 · p. 10

5 CONCLUSION

• The conclusion presents VICS-G as a practical, training-free decoder for frozen VLA policies that reduces observed safety cost while retaining near-policy completion and sustained progress.It also identifies execution feedback and simulator lookahead as extensions to richer deployment settings.Source: §5 · p. 10

A QUALITATIVE EPISODE COMPARISONS

• In both selected episodes, VICS-G completes the task at lower cost than policy sampling, while RCD’s lower cost accompanies unsuccessful termination.These qualitative cases complement the aggregate results reported elsewhere in the paper.Source: A QUALITATIVE EPISODE COMPARISONS · p. 13

A.1 NAVIGATION: RECOVERING FROM FAILED RETRIES

• In ObjectNav, VICS-G replaces failed retries and completes at cost 11; policy sampling ends unsuccessfully at cost 23, while RCD stops early at cost 3.VICS-G leaves the failure region and continues after its cost levels off.Source: Figure 3 · p. 13, §A.1 · p. 13

A.2 FETCH: COMPLETING THE INTENDED GRASP

• At cost 12, VICS-G finishes holding the target plant, versus policy sampling’s wrong object at cost 19 and RCD ending empty-handed at cost 3.The later frames are aligned by interaction phase rather than elapsed time.Source: Figure 4 · p. 14, §A.2 · p. 14

B HISTORY-CONDITIONED ROBOT DECODING

• The history-conditioned trajectory law combines the frozen policy, environment dynamics, and observation process to represent possible futures from the observable history.The model distinguishes latent physical state from the history available to the policy.Source: B HISTORY-CONDITIONED ROBOT DECODING · p. 14, §B.1 · p. 15

B.1 REFERENCE TRAJECTORY TARGET

• With bounded total reward and cost and positive prior probability of safe completion, the prior-preserving feasible target has a unique solution; positive reward-cost weighting changes feasible mass without changing its support.When β = λ = 0, feasible-future mass equals the probability of safe completion.Source: §B.1 · p. 15

C.1 PROOF OF THE PRIOR-PRESERVING FEASIBLE TARGET

• The proof obtains the target by restricting the prior to safe-completion trajectories, applying an exponential reward-cost tilt, and normalizing the resulting distribution.The KL term’s strict convexity on the feasible set establishes uniqueness up to prior-null trajectories.Source: §C.1 · p. 16

C.2 PROOF OF THE FEASIBLE-FUTURE IDENTITY

• The ideal next-block score adds the policy log-likelihood and local reward-cost terms to the log feasible-future mass, whose suffix law integrates latent-state and transition uncertainty.A zero feasible-future mass gives a score of negative infinity.Source: §C.2 · p. 16

C.3 PROOF OF THE PARTIAL-SUPPORT FINITE-CANDIDATE THEOREM

• Under score separation from surviving dead ends and bounded pairwise score error among viable survivors, the selected candidate is viable and its regret within the retained viable set is at most ε.The result also bounds regret relative to the comparison set by the proposal/admission loss plus ε.Source: §C.3 · pp. 16–17
• The regret bound separates proposal or admission loss from ranking error, while complete rejection yields no survivors and no support-separation requirement.The theorem’s finite-candidate guarantee applies at a fixed history.Source: §C.3 · p. 17
• On the authorized reranking branch, the bound applies under its assumptions; when reranking is not invoked, the policy sample can have infinite loss if its feasible-future mass is zero.The policy sample is included in the viable comparison set whenever it is viable.Source: §C.3 · p. 17

C.4 SAFE-DEAD-END SEPARATION

• Two candidates can share likelihood and local reward-cost terms yet differ in feasible-future mass, so the ideal rule excludes the candidate with no safe policy-supported continuation.The viable candidate has positive feasible-future mass; the dead end has zero mass.Source: §C.4 · pp. 17–18

D SCORE ERROR AND ROLLOUT RECOVERY

• VICS-G’s generic score combines policy log-likelihood, local robustness, and a continuation correction; experiments compare outcomes but do not estimate the latent score errors.Both the local robustness term and continuation correction contribute to ranking error.Source: D SCORE ERROR AND ROLLOUT RECOVERY · p. 18
• With bounded rollout weights and a positive feasible-mass floor for viable candidates, sufficiently many continuation samples preserve the unique ideal-score maximizer with probability at least 1 − δ.Zero-support candidates have estimated continuation mass zero almost surely; near ties and rare safe completions require more rollouts.Source: D SCORE ERROR AND ROLLOUT RECOVERY · pp. 18–19

E EXACT FINITE-MODEL EXAMPLES

• The four-action example makes the policy MAP action locally unsafe, local filtering choose a dead end, and exact support plus magnitude choose the higher-valued feasible action.The feasible-action odds reverse from 1/2 to 2 after magnitude weighting; accepting the dead end at the initial gate bypasses later checks.Source: E EXACT FINITE-MODEL EXAMPLES · p. 19
• With identical likelihoods and local terms, two worlds require opposite actions for safe completion, so a local-only decoder averages 1/2 success while the feasible-future rule succeeds in both.The example establishes that continuation-dependent information is necessary for uniform success on this pair of worlds.Source: E EXACT FINITE-MODEL EXAMPLES · p. 19
• In the three-action example, both viable actions succeed surely but differ in suffix reward; exact support avoids the dead end, while exact magnitude selects the higher-valued continuation.Regret is measured in base-two score units.Source: E EXACT FINITE-MODEL EXAMPLES · pp. 19–20, Table 3 · p. 20
• Finite candidate subsets distinguish missing viable proposals, false rejection of viable candidates, and misranking survivors; with pairwise total-score error at most ε, within-set regret is at most ε.Sound rejection eliminates false-rejection loss, while complete rejection removes unresolved dead ends.Source: E EXACT FINITE-MODEL EXAMPLES · p. 20

F DECODING RULES AND EVALUATION PROTOCOL

• Across six Safety-CHORES settings, the protocol evaluates frozen base and safety-aligned policies on shared episode sets and reports success, cumulative cost, violations, and episode length jointly.The generic VICS-G coefficients are shared across settings; coefficients were selected using approximately 10% of episodes per task family.Source: F DECODING RULES AND EVALUATION PROTOCOL · p. 20

F.1 GENERIC SCORE AND EXECUTION RULE

• The decoder’s local score penalizes observed harm, repeated actions, and failed retries, while zG adds continuation preference and a repetition term without score normalization or clipping.The evaluation assumes monitor signals beyond images, including executed actions, failures, collisions, simulator costs, and object visibility and distance.Source: §F.1 · pp. 20–21
• VICS forms candidates from the policy sample and highest-likelihood actions, then gates intervention and filters by trust region, predicted safety, and task feasibility before reranking.The sample is retained if the gate accepts it, no candidate survives admission, or replacement authorization fails.Source: §F.1 · pp. 21–22
• PickUp and ObjectNav symbolic variants add task-grounded eligibility rules, while Fetch uses a rule-filtered recovery procedure over all 20 actions; these rules do not certify complete viability.When no symbolic restriction applies, generic and symbolic selection agree under the same candidate order and scores.Source: §F.1 · p. 22
• RCD scores policy likelihood with local robustness but predicts no future trajectory and uses no VICS gate, trust-region filter, symbolic support rule, or replacement veto.The ηv = 0 ablation retains VICS-G’s gate, filters, history-dependent local score, trust region, and fallback, isolating the additional zG correction.Source: §F.1 · pp. 22–23

G FEASIBILITY-WEIGHT SENSITIVITY

• The matched sweeps vary ηv over 0, 0.05, 0.10, and 0.20 across both checkpoints and all three task settings, with complete episode coverage at each weight.Fetch uses all 20 actions in the sweep, whereas the main Fetch comparisons use top-12 plus the policy sample.Source: G FEASIBILITY-WEIGHT SENSITIVITY · p. 23
• The 95% pointwise bootstrap intervals assume exchangeable episodes conditional on checkpoint and omit scene clustering, retraining or repeated-execution variability, and multiplicity correction.A zero-width interval from identical observed success outcomes does not establish population equivalence.Source: G FEASIBILITY-WEIGHT SENSITIVITY · p. 23
• At ηv = 0.10, mean violations decrease in five of six settings and mean cost in four, but all cost and violation intervals include zero.Table 5 reports the incremental effect relative to zero weight; the gate and history-dependent local score remain active in both arms.Source: Table 5 · p. 24, G FEASIBILITY-WEIGHT SENSITIVITY · p. 24
• ObjectNav-base success falls by 3.00 percentage points at ηv = 0.10, while cost, violations, and episode length increase; ObjectNav-safe also loses success.At 0.20, ObjectNav-base improves relative to 0.10 but remains worse than at zero weight.Source: G FEASIBILITY-WEIGHT SENSITIVITY · p. 24

H ROLLOUT-AUGMENTED VICS-G: PROTOCOL AND RESULTS

• The comparison uses all 172 Fetch tasks, matched seeds, and a modified Unity/PhysX simulator, with each method evaluated once per task.Replay checks establish agreement in observed branching conditions and unchanged-action continuations, but cannot certify exact restoration of hidden simulator state.Source: H ROLLOUT-AUGMENTED VICS-G: PROTOCOL… · p. 26
• VICS-R activates once during the first 16 steps, comparing the current action with up to two alternatives using four matched continuations per candidate before returning control to VICS-G.Hard safety and terminal constraints remain active throughout.Source: H ROLLOUT-AUGMENTED VICS-G: PROTOCOL… · p. 26
• On safety-aligned Fetch, VICS-R leaves success unchanged at 58.72%, raises safe success from 27.91% to 30.23%, and reduces mean cost by 3.6% versus VICS-G.The paired safe-success difference is not statistically significant (two-sided exact McNemar p = 0.344).Source: Table 6 · p. 26, H ROLLOUT-AUGMENTED VICS-G: PROTOCOL… · p. 26
• The rollout comparison adds both continuation evidence and an intervention opportunity, so it does not isolate the additive correction; four samples do not establish the stated corollary’s assumptions.Larger budgets and repeated lookahead are identified as future work.Source: H ROLLOUT-AUGMENTED VICS-G: PROTOCOL… · p. 26

I EXECUTION FEEDBACK: RESULTS AND STATISTICAL EVIDENCE

• The execution-feedback study compares policy sampling, RCD, switch-after-failure, and VICS-G with and without feedback under clean and perturbed conditions.The supplied passage describes the comparison design, not its outcomes.Source: I EXECUTION FEEDBACK: RESULTS AND… · p. 27

I.1 DESIGN AND PAIRED ANALYSIS

• The prespecified study paired methods across 172 tasks and perturbations; supplementary clean and post-failure comparisons expanded the complete study to 4,300 executions.Two-sided 95% percentile intervals used 10,000 bootstrap resamples of 172 paired task differences, with perturbations matched across methods.Source: §I.1 · pp. 27–28

I.2 COMPLETION–SAFETY TRADE-OFFS ACROSS CONDITIONS

• Under both shifts, the feedback-aware decoder matched or exceeded policy-sampling success while lowering observed cost and violations at similar episode lengths; RCD had lower cost but substantially lower completion.These are descriptive operating points across clean, IID, and persistent conditions.Source: §I.2 · p. 28, Table 7 · p. 28

I.3 INCREMENTAL FEEDBACK EFFECT BY CONDITION

• Feedback lowered mean cost and violations versus masking under both perturbations, but not without perturbations; all paired intervals included zero, leaving the condition-dependent pattern inconclusive.The condition-specific comparisons are exploratory, and the clean comparison was added after initial results.Source: §I.3 · p. 29, Table 8 · p. 29

I.4 CONCLUSIONS FROM THE PRESPECIFIED TESTS

• Feedback met the 5-percentage-point success tolerance versus masking (+0.58 pp), but the cost comparison was inconclusive, so the prespecified testing sequence stopped.Supplementary comparisons with policy sampling remain descriptive.Source: Table 9 · p. 29, §I.4 · p. 29

B. Descriptive comparisons: reference is policy sampling

• The policy-sampling comparison reports a descriptive safety-cost difference of −0.688; it was not tested because the sequence stopped at Step 2.Supplementary comparisons are excluded from the prespecified tests.Source: B. Descriptive comparisons: reference… · p. 29

J RELATED WORK

• The trajectory tilt draws on KL-regularized control (Levine, 2018), while the target tracks policy-supported safe-completion suffixes and the deployed score is a selective heuristic.The paper distinguishes its policy-relative target from prior inference-time guidance and future-aware decoding approaches (Yang & Klein, 2021; Lu et al., 2022; Ji et al., 2025; Kapoor et al., 2025).Source: J RELATED WORK · p. 30
• The support notion is narrower than viability over all admissible controls: the decoder uses the frozen continuation law and claims neither invariant-set nor closed-loop guarantees.The evaluation also treats completion and trajectory safety jointly rather than reading lower unconditioned cost alone as dominance.Source: J RELATED WORK · p. 30